2507.07907
Model: healer-alpha-free
# A statistical physics framework for optimal learning
**Authors**:
- Francesca Mignacco, Francesco Mori (Graduate Center, City University of New York, New York, NY 10016, USA)
## Abstract
Learning is a complex dynamical process shaped by a range of interconnected decisions. Careful design of hyperparameter schedules for artificial neural networks or efficient allocation of cognitive resources by biological learners can dramatically affect performance. Yet, theoretical understanding of optimal learning strategies remains sparse, especially due to the intricate interplay between evolving meta-parameters and nonlinear learning dynamics. The search for optimal protocols is further hindered by the high dimensionality of the learning space, often resulting in predominantly heuristic, difficult to interpret, and computationally demanding solutions. Here, we combine statistical physics with control theory in a unified theoretical framework to identify optimal protocols in prototypical neural network models. In the high-dimensional limit, we derive closed-form ordinary differential equations that track online stochastic gradient descent through low-dimensional order parameters. We formulate the design of learning protocols as an optimal control problem directly on the dynamics of the order parameters with the goal of minimizing the generalization error at the end of training. This framework encompasses a variety of learning scenarios, optimization constraints, and control budgets. We apply it to representative cases, including optimal curricula, adaptive dropout regularization and noise schedules in denoising autoencoders. We find nontrivial yet interpretable strategies highlighting how optimal protocols mediate crucial learning tradeoffs, such as maximizing alignment with informative input directions while minimizing noise fitting. Finally, we show how to apply our framework to real datasets. Our results establish a principled foundation for understanding and designing optimal learning protocols and suggest a path toward a theory of meta-learning grounded in statistical physics.
## 1 Introduction
Learning is intrinsically a multilevel process. In both biological and artificial systems, this process is defined through a web of design choices that can steer the learning trajectory toward crucially different outcomes. In machine learning (ML), this multilevel structure underlies the optimization pipeline: model parameters are adjusted by a learning algorithmâe.g., stochastic gradient descent (SGD)âthat itself depends on a set of higherâorder decisions, specifying the network architecture, hyperparameters, and dataâselection procedures [1]. These meta-parameters are often adjusted dynamically throughout training following predefined schedules to enhance performance. Biological learning is also mediated by a range of control signals across scales. Cognitive control mechanisms are known to modulate attention and regulate learning efforts to improve flexibility and multi-tasking [2, 3, 4]. Additionally, structured training protocols are widely adopted in animal and human training to make learning processes faster and more robust. For instance, curricula that progressively increase the difficulty of the task often improve the final performance [5, 6].
Optimizing the training schedulesâeffectively âlearning to learnââis a crucial problem in ML. However, the proposed solutions remain largely based on trial-and-error heuristics and often lack a principled assessment of their optimality. The increasing complexity of modern ML architectures has led to a proliferation of meta-parameters, exacerbating this issue. As a result, several paradigms for automatic learning, such as meta-learning and hyperparameter optimization [7, 8], have been developed. Proposed methods range from grid and random hyperparameter searches [9] to Bayesian approaches [10] and gradientâbased metaâoptimization [11, 12]. However, these methods operate in highâdimensional, nonconvex search spaces, making them computationally expensive and often yielding strategies that are hard to interpret. Although one can frame the selection of training protocols as an optimalâcontrol (OC) problem, applying standard control techniques to the full parameter space is often infeasible due to the curse of dimensionality.
Statistical physics provides a long-standing theoretical framework for understanding learning through prototypical models [13], a perspective that has carried over into recent advances in ML theory [14, 15]. It exploits the high dimensionality of learning problems to extract low-dimensional effective descriptions in terms of order parameters that capture the key properties of training and performance. A substantial body of theoretical results has been obtained in the Bayes-optimal setting, characterizing the information-theoretically optimal performance for given data-generating processes and providing a threshold that no algorithm can improve [16, 17]. In parallel, the algorithmic performance of practical procedures, such as empirical risk minimization, has been studied both in the asymptotic regime via equilibrium statistical mechanics [18, 19, 20, 21, 22, 23] and through explicit analyses of training dynamics [24, 25, 26, 27, 28]. More recently, neural network models analyzed with statistical physics methods have been used to study various paradigmatic learning settings relevant to cognitive science [29, 30, 31]. However, these lines of work have mainly focused on predefined protocols, often keeping meta-parameters constant during training, without addressing the derivation of optimal learning schedules.
In this paper, we propose a unified framework for optimal learning that combines statistical physics and control theory to systematically identify training schedules across a broad range of learning scenarios. Specifically, we define an OC problem directly on the low-dimensional dynamics of the order parameters, where the meta-parameters of the learning process serve as controls and the final performance is the objective. This approach serves as a testbed for uncovering general principles of optimal learning and offers two key advantages. First, the reduced descriptions of the learning dynamics circumvent the curse of dimensionality, enabling the application of standard control-theoretic techniques. Second, the order parameters capture essential aspects of the learning dynamics, allowing for a more interpretable analysis of why the resulting strategies are effective.
In particular, we consider online training with SGD in a general two-layer network model that includes several learning settings as special cases. Building on the foundational work of [32, 33, 34], we derive exact closed-form equations describing the evolution of the relevant order parameters during training. Control-theoretical techniques can then be applied to identify optimal training schedules that maximize the final performance. This formulation enables a unified treatment of diverse learning paradigms and their associated meta-parameter schedules, such as task ordering, learning rate tuning, and dynamic modulation of the node activations. A variety of learning constraints and control budgets can be directly incorporated. Our work contributes to the broader effort to develop theoretical frameworks for the control of nonequilibrium systems [35, 36, 37], given that learning dynamics are high-dimensional, stochastic, and inherently nonequilibrium processes.
While we present our approach here in full generality, a preliminary application of this method for optimal task-ordering protocols in continual learning was recently presented in the conference paper [38]. Related variational approaches were explored in earlier work from the 1990s, primarily in the context of learning rate schedules [39, 40]. More recently, computationally tractable meta-learning strategies have been studied in linear networks [41, 42]. However, a general theoretical framework for identifying optimal training protocols in nonlinear networks is still missing.
The rest of the paper is organized as follows. In Section 2, we introduce the theoretical framework. Specifically, we present the model in Section 2.1 and we define the order parameters and derive the dynamical equations for online SGD training in Section 2.2. The control-theoretic techniques used throughout the paper are described in Section 2.3. In Section 2.4, we illustrate a range of learning scenarios that can be addressed within this framework. In Section 3, we derive and discuss optimal training schedules in three representative settings: curriculum learning (Section 3.1), dropout regularization (Section 3.2), and denoising autoencoders (Section 3.3). We conclude in Section 4 with a summary of our findings and a discussion of open directions. Additional technical details are provided in the appendices.
## 2 Theoretical framework
### 2.1 The model
We study a general learning framework based on the sequence multi-index model introduced in [43]. This model captures a broad class of learning scenarios, both supervised and unsupervised, and admits a closed-form analytical description of its training dynamics. This dual feature allows us to derive optimal learning strategies across various regimes and to highlight multiple potential applications. We begin by presenting a general formulation of the model, followed by several concrete examples.
We consider a dataset $\mathcal{D}=\bigl{\{}(\bm{x}^{\mu},y^{\mu})\bigr{\}}_{\mu=1}^{P}$ of $P$ samples, where $\bm{x}^{\mu}\in\mathbb{R}^{N\times L}$ are i.i.d. inputs and $y^{\mu}\in\mathbb{R}$ are the corresponding labels (if supervised learning is considered). Each input sample ${\bm{x}}\in\mathbb{R}^{N\times L}$ , a sequence with $L$ elements ${\bm{x}}_{l}$ of dimension $N$ , is drawn from a Gaussian mixture
$$
{\bm{x}}_{l}\sim\mathcal{N}\left(\frac{{\bm{\mu}}_{l,c_{l}}}{\sqrt{N}},\sigma^
{2}_{l,c_{l}}\bm{I}_{N}\right)\,, \tag{1}
$$
where $c_{l}\in\{1\,,\ldots\,,C_{l}\}$ denotes cluster membership. The random vector ${\bm{c}}=\{c_{l}\}_{l=1}^{L}$ is sampled from a probability distribution $p_{c}({\bm{c}})$ , which can encode arbitrary correlations. In supervised settings, we will often assume
$$
y=f^{*}_{{\bm{w}}_{*}}({\bm{x}})+\sigma_{n}z,\qquad z\sim\mathcal{N}(0,1), \tag{2}
$$
where $f^{*}_{{\bm{w}}_{*}}({\bm{x}})$ is a fixed teacher network with $M$ hidden units and parameters ${\bm{w}}_{*}\in\mathbb{R}^{N\times M}$ , and $\sigma_{n}$ controls label noise. This teacherâstudent (TS) paradigm is standard in statistical physics and it allows for analytical characterization [44, 45, 32, 33, 34, 13, 24].
We consider a two-layer neural network $f_{\bm{w},\bm{v}}(\bm{x})=\tilde{f}\bigl{(}\tfrac{\bm{x}^{\top}\,\bm{w}}{\sqrt {N}},\mathbf{v}\bigr{)}$ with $K$ hidden units. In a TS setting, this network serves as the student. The parameters $\bm{w}\in\mathbb{R}^{N\times K}$ (first-layer) and $\bm{v}\in\mathbb{R}^{K\times H}$ (readout) are both trainable. The readout $\bm{v}$ has $H$ heads, $\bm{v}_{h}\in\mathbb{R}^{K}$ for $h=1,\dots,H$ , which can be switched to adapt to different contexts or tasks. In the simplest case, $H=L=1$ , the network will often take the form
$$
f_{\bm{w},\bm{v}}(\bm{x})=\frac{1}{\sqrt{K}}\sum_{k=1}^{K}v_{k}\leavevmode
\nobreak\ g\left(\frac{{\bm{w}}_{k}\cdot{\bm{x}}}{\sqrt{N}}\right)\,, \tag{3}
$$
where we have dropped the head index, and $g(\cdot)$ is a nonlinearity (e.g., $g(z)=\operatorname{erf}(z/\sqrt{2}))$ .
To characterize the learning process, we consider a cost function of the form
$$
\mathcal{L}({\bm{w}},{\bm{v}}|\bm{x},\bm{c})=\ell\left(\frac{{\bm{x}}^{\top}{
\bm{w}_{*}}}{\sqrt{N}},\frac{{\bm{x}}^{\top}{\bm{w}}}{\sqrt{N}},\frac{\bm{w}^{
\top}\bm{w}}{N},{\bm{v}},{\bm{c}},z\right)+\tilde{g}\left(\frac{\bm{w}^{\top}
\bm{w}}{N},{\bm{v}}\right)\,, \tag{4}
$$
where we have introduced the loss function $\ell$ , and the regularization function $\tilde{g}$ , which typically penalizes large values of the parameter norms. Note that the functional form of $\ell(\cdot)$ in Eq. (4) implicitly contains details of the problem, including the network architecture, the specific loss function used, and the shape of the target function. Additionally, it may contain adaptive hyperparameters and controls on architectural features. When considering a TS setting, the loss takes the form
$$
\ell\left(\frac{{\bm{x}}^{\top}{\bm{w}_{*}}}{\sqrt{N}},\frac{{\bm{x}}^{\top}{
\bm{w}}}{\sqrt{N}},\frac{\bm{w}^{\top}\bm{w}}{N},{\bm{v}},{\bm{c}},z\right)=
\tilde{\ell}(f_{\bm{w},\bm{v}}(\bm{x}),y)\,, \tag{5}
$$
where $y$ is given in Eq. (2) and $\tilde{\ell}(a,b)$ penalizes dissimilar values of $a$ and $b$ . A typical choice is the square loss: $\tilde{\ell}(a,b)=(a-b)^{2}/2$ .
### 2.2 Learning dynamics
We study the learning dynamics under online (oneâpass) SGD, in which each update is computed using a fresh sample $\bm{x}^{\mu}$ at each training step $\mu$ In contrast, offline (multi-pass) SGD repeatedly reuses the same samples throughout training.. This regime admits an exact analysis via statisticalâphysics methods [32, 33, 34, 24]. The parameters evolve as
$$
\displaystyle{\bm{w}}^{\mu+1}={\bm{w}}^{\mu}-{\eta}\nabla_{\bm{w}}\mathcal{L}(
{\bm{w}}^{\mu},{\bm{v}}^{\mu}|\bm{x}^{\mu},\bm{c}^{\mu})\;, \displaystyle\bm{v}^{\mu+1}=\bm{v}^{\mu}-\frac{\eta_{v}}{N}\nabla_{\bm{v}}
\mathcal{L}({\bm{w}}^{\mu},{\bm{v}}^{\mu}|\bm{x}^{\mu},\bm{c}^{\mu})\;, \tag{6}
$$
where $\eta$ and $\eta_{v}$ denote the learning rates of the first-layer and readout parameters. Other training algorithms, such as biologically plausible learning rules [46, 47], can be incorporated into this framework, but we leave their analysis to future work. We focus on the high-dimensional limit where the dimensionality of the input layer $N$ and the number of training epochs $\mu$ , jointly tend to infinity at fixed training time $\alpha=\mu/N$ . All other dimensions, i.e., $K$ , $H$ , $L$ and $M$ , are assumed to be $\mathcal{O}_{N}(1)$ .
The generalization error is given by
$$
\epsilon_{g}({\bm{w}},{\bm{v}})=\mathbb{E}_{\bm{x},\bm{c}}\left[\ell_{g}\left(
\frac{{\bm{x}}^{\top}{\bm{w}_{*}}}{\sqrt{N}},\frac{{\bm{x}}^{\top}{\bm{w}}}{
\sqrt{N}},\frac{\bm{w}^{\top}\bm{w}}{N},{\bm{v}},{\bm{c}},0\right)\right]\,, \tag{7}
$$
where $\mathbb{E}_{\bm{x},\bm{c}}$ denotes the expectation over the joint distribution of $\bm{x}$ and ${\bm{c}}$ and the label noise $z$ is set to zero. Depending on the context, the function $\ell_{g}$ may coincide with the training loss $\ell$ , or it may represent a different metricâsuch as the misclassification error in the case of binary labels. Crucially, the generalization error $\epsilon_{g}({\bm{w}},{\bm{v}})$ depends on the high-dimensional first-layer weights only through the following low-dimensional order parameters:
$$
Q^{\mu}_{kk^{\prime}}\coloneqq\frac{{\bm{w}^{\mu}_{k}}\cdot\bm{w}^{\mu}_{k^{
\prime}}}{N}\;,\quad M^{\mu}_{km}\coloneqq\frac{{\bm{w}^{\mu}_{k}}\cdot\bm{w}_
{*,m}}{N}\;,\quad R^{\mu}_{k(l,c_{l})}\coloneqq\frac{{\bm{w}^{\mu}_{k}}\cdot
\bm{\mu}_{l,c_{l}}}{{N}}\;. \tag{8}
$$
Collecting these together with the readout parameters $\bm{v}^{\mu}$ into a single vector
$$
\mathbb{Q}=\left({\rm vec}\left({\bm{Q}}\right),{\rm vec}\left({\bm{M}}\right)
,{\rm vec}\left({\bm{R}}\right),{\rm vec}\left({\bm{v}}\right)\right)^{\top}
\in\mathbb{R}^{K^{2}+KM+K(C_{1}+\ldots+C_{L})+HK}\,, \tag{9}
$$
we can write $\epsilon_{g}({\bm{w}},{\bm{v}})=\epsilon_{g}(\mathbb{Q})$ (see Appendix A). Additionally, it is useful to define the low-dimensional constant parameters
$$
\displaystyle\begin{split}S_{m(l,c_{l})}\coloneqq\frac{{\bm{w}_{*,m}}\cdot\bm{
\mu}_{l,c_{l}}}{{N}}\;,\quad T_{mm^{\prime}}\coloneqq\frac{{\bm{w}_{*,m}}\cdot
\bm{w}_{*,m^{\prime}}}{N}\;,\quad\Omega_{(l,c_{l})(l^{\prime},c^{\prime}_{l^{
\prime}})}=\frac{\bm{\mu}_{l,c_{l}}\cdot\bm{\mu}_{l^{\prime},c^{\prime}_{l^{
\prime}}}}{N}\;.\end{split} \tag{10}
$$
Note that the scaling of teacher vectors $\bm{w}_{*,m}$ and the centroids $\bm{\mu}_{l,c_{l}}$ with $N$ is chosen so that the parameters in Eq. (10) are $\mathcal{O}_{N}(1)$ .
In the highâdimensional limit, the stochastic fluctuations of the order parameters $\mathbb{Q}$ vanish and their dynamics concentrate on a deterministic trajectory. Consequently, $\mathbb{Q}(\alpha)$ satisfies a closed system of ordinary differential equations (ODEs) [32, 33, 34, 13, 24]:
$$
\displaystyle\frac{{\rm d}\mathbb{Q}}{{\rm d}\alpha}=f_{\mathbb{Q}}\left(
\mathbb{Q}(\alpha),\bm{u}(\alpha)\right)\;,\qquad{\rm with}\quad\alpha\in(0,
\alpha_{F}]\;, \tag{11}
$$
where $\alpha_{F}=P/N$ denotes the final training time and the explicit form of $f_{\mathbb{Q}}$ is provided in Appendix A. In Appendix C, we check these theoretical ODEs via numerical simulations, finding excellent agreement. The vector $\bm{u}(\alpha)$ encodes controllable parameters involved in the training process. We assume that ${\bm{u}}(\alpha)\in\mathcal{U}$ , where $\mathcal{U}$ is the set of feasible controls, whose dimension is $\mathcal{O}_{N}(1)$ . The set $\mathcal{U}$ may include discrete, continuous, or mixed controls. For example, setting $\bm{u}(\alpha)=\eta(\alpha)$ corresponds to dynamic learningârate schedules. The control $\bm{u}(\alpha)$ could also parameterize a time-dependent distribution of the cluster variable $\bm{c}$ to encode sample difficulty, e.g., to study curriculum learning. Likewise, $\bm{u}(\alpha)$ could describe aspects of the network architecture, e.g., a timeâdependent dropout rate. Several specific examples are discussed in Section 2.4.
Identifying optimal schedules for $\bm{u}(\alpha)$ is the central goal of this work. Solving this control problem directly in the original highâdimensional parameter space is computationally challenging. However, the exact lowâdimensional description of the training dynamics in Eq. (11) allows to readily apply standard OC techniques.
### 2.3 Optimal control of the learning dynamics
In this section, we describe the OC framework that allows us to identify optimal learning strategies. We seek to identify the OC $\bm{u}(\alpha)\in\mathcal{U}$ that minimizes the generalization error at the end of training, i.e., at training time $\alpha_{F}$ . To this end, we introduce the cost functional
$$
\mathcal{F}[\bm{u}]=\epsilon_{g}(\mathbb{Q}(\alpha_{F}))\,, \tag{12}
$$
where the square brackets indicate functional dependence on the full control trajectory $\bm{u}(\alpha)$ , for $0\leq\alpha\leq\alpha_{F}$ . The functional dependence on $\bm{u}(\alpha)$ appears implicitly through the ODEs (11), which govern the evolution from the fixed initial state $\mathbb{Q}(0)=\mathbb{Q}_{0}$ to the final state $\mathbb{Q}(\alpha_{F})$ . Note that, while we consider globally optimal schedulesâthat is, schedules optimized with respect to the final cost functionalâprevious works have also explored greedy schedules that are locally optimal, maximizing the error decrease or the learning speed at each training step [48, 49]. These schedules are easier to analyze but generally lead to suboptimal results [40]. Furthermore, although our focus is on minimizing the final generalization error, the framework can accommodate alternative objectives. For instance, one may optimize the timeâaveraged generalization error as in [41], if the performance during training, rather than only at $\alpha_{F}$ , is of interest. We adopt two types of OC techniques: indirect methods, which solve the boundaryâvalue problem defined by the Pontryagin maximum principle [50, 51, 52], and direct methods, which discretize the control $\bm{u}(\alpha)$ and map the problem into a finiteâdimensional nonlinear program [53]. Additional costs or constraints associated with the control signal ${\bm{u}}$ can be directly incorporated into both classes of methods.
#### 2.3.1 Indirect methods
Following Pontryaginâs maximum principle [50], we augment the functional in Eq. (12) by introducing the Lagrange multipliers $\hat{\mathbb{Q}}(\alpha)$ to enforce the dynamics (11)
$$
\mathcal{F}[\bm{u},\mathbb{Q},\hat{\mathbb{Q}}]=\epsilon_{g}\bigl{(}\mathbb{Q}
(\alpha_{F})\bigr{)}+\int_{0}^{\alpha_{F}}{\rm d}\alpha\;\hat{\mathbb{Q}}(
\alpha)\cdot\left[-\frac{{\rm d}\mathbb{Q}(\alpha)}{{\rm d}\alpha}+f_{\mathbb{
Q}}\bigl{(}\mathbb{Q}(\alpha),\,\bm{u}(\alpha)\bigr{)}\right], \tag{13}
$$
where $\hat{\mathbb{Q}}(\alpha)$ are known as adjoint (or costate) variables. The optimality conditions are $\delta\mathcal{F}/\delta\hat{\mathbb{Q}}(\alpha)=0$ and $\delta\mathcal{F}/\delta\mathbb{Q}(\alpha)=0$ . The first yields the forward dynamics (11). For $\alpha<\alpha_{F}$ , the second, after integration by parts, gives the adjoint (backward) ODEs
$$
\displaystyle-\frac{{\rm d}\hat{\mathbb{Q}}(\alpha)^{\top}}{{\rm d}\alpha} \displaystyle=\hat{\mathbb{Q}}(\alpha)^{\top}\nabla_{\mathbb{Q}}f_{\mathbb{Q}}
\bigl{(}\mathbb{Q}(\alpha),\bm{u}(\alpha)\bigr{)}, \tag{14}
$$
with the final condition at $\alpha=\alpha_{F}$ :
$$
\hat{\mathbb{Q}}(\alpha_{F})=\nabla_{\mathbb{Q}}\,\epsilon_{g}\bigl{(}\mathbb{
Q}(\alpha_{F})\bigr{)}. \tag{15}
$$
Variations at $\alpha=0$ are not considered since $\mathbb{Q}(0)=\mathbb{Q}_{0}$ is fixed. Finally, optimizing $\bm{u}$ point-wise yields
$$
\bm{u}^{*}(\alpha)=\underset{\bm{u}\in\mathcal{U}}{\arg\min}\;\bigl{\{}\hat{
\mathbb{Q}}(\alpha)\cdot\,f_{\mathbb{Q}}\bigl{(}\mathbb{Q}(\alpha),\bm{u}\bigr
{)}\bigr{\}}. \tag{16}
$$
In practice, we use the forward-backward sweep method: starting from an initial guess for $\bm{u}$ , we iterate the following steps until convergence.
1. Integrate $\mathbb{Q}$ forward via (11) from $\mathbb{Q}(0)=\mathbb{Q}_{0}$ .
1. Integrate $\hat{\mathbb{Q}}$ backward via (14) from $\hat{\mathbb{Q}}(\alpha_{F})$ in (15).
1. Update $\bm{u}^{k+1}(\alpha)=\gamma_{\rm damp}\bm{u}^{k}(\alpha)+(1-\gamma_{\rm damp}) \bm{u}^{*}(\alpha)$ , where $\bm{u}^{*}(\alpha)$ is given in (16).
We typically choose the damping parameter $\gamma_{\rm damp}>0.9$ . Convergence is usually reached within a few hundred to a few thousand iterations.
#### 2.3.2 Direct methods
Direct methods discretize the control trajectory $\bm{u}(\alpha)$ on a finite grid of $I=\alpha_{F}/{\rm d}\alpha$ intervals and map the continuousâtime OC problem into a finiteâdimensional nonlinear program (NLP). We introduce optimization variables for $\mathbb{Q}$ and $\bm{u}$ at each node $\alpha_{j}=j\leavevmode\nobreak\ {\rm d}\alpha$ , enforce the dynamics (11) via constraints on each interval, and solve the resulting NLP using the CasADi package [54]. In this paper, we implement a multipleâshooting scheme: $\bm{u}(\alpha)$ is parameterized as constant on each interval, and continuity of $\mathbb{Q}$ is enforced at the boundaries. While direct methods are conceptually simplerârelying on standard NLP solvers and avoiding the explicit derivation of adjoint equationsâin the settings under consideration, we find that they tend to perform worse when the control $\bm{u}$ has discrete components. Conversely, indirect methods require computing costate derivatives but yield more accurate solutions for discrete controls. Depending on the problem setting, we therefore choose between direct and indirect approaches as specified in each case.
### 2.4 Special cases of interest
In this section, we illustrate how the proposed framework can be readily applied to describe several representative learning scenarios, addressing theoretical questions emerging in machine learning and cognitive science. We organize the presentation of different learning strategies into three main categories, each reflecting a distinct aspect of the training process: hyperparameters of the optimization, data selection mechanisms, and architectural adaptations.
#### 2.4.1 Hyperparameter schedules
Optimization hyperparameters are external configuration variables that shape the dynamics of the learning process. Dynamically tuning these parameters during training is a standard practice in machine learning, and represents one of the most widely used and studied forms of training protocols.
Learning rate.
The learning rate $\eta$ is often regarded as the single most important hyperparameter [1]. A small $\eta$ mitigates the impact of data noise but slows convergence, whereas a large $\eta$ accelerates convergence at the expense of amplified stochastic fluctuations, which can lead to divergence of the training dynamics. Consequently, many empirical studies have proposed heuristic schedules, such as initial warmâups [55] or periodic schemes [56], and methods to optimize $\eta$ via additional gradient steps [57]. From a theoretical perspective, optimal learning rate schedules were already investigated in the 1990s in the context of online training of two-layer networks, using a variational approach closely related to ours [39, 40, 58]. More recently, [59] analytically derived optimal learning rate schedules to optimize high-dimensional non-convex landscapes. Within our framework, the learning rate can be always included in the control vector $\bm{u}$ , as done in [38] focusing on online continual learning. Optimal learning rate schedules are further discussed in the context of curriculum learning in Section 3.1.
Batch size.
Dynamically adjusting the batch size, i.e., the number of data samples used to estimate the gradient at each SGD step, has been proposed as a powerful alternative to learning rate schedules [60, 61, 62]. Mini-batch SGD can be treated within our theoretical formulation by identifying the batch of samples with the input sequence, corresponding to a loss function of the form:
$$
\displaystyle\ell\left(\frac{{\bm{x}}^{\top}{\bm{w}_{*}}}{\sqrt{N}},\frac{{\bm
{x}}^{\top}{\bm{w}}}{\sqrt{N}},\frac{\bm{w}^{\top}\bm{w}}{N},{\bm{v}},{\bm{c}}
,z\right)=\frac{1}{L}\sum_{l=1}^{L}\hat{\ell}\left(\frac{{\bm{w}_{*}}^{\top}{
\bm{x}}_{l}}{\sqrt{N}},\frac{{\bm{w}}^{\top}{\bm{x}}_{l}}{\sqrt{N}},\frac{\bm{
w}^{\top}\bm{w}}{N},{\bm{v}},c_{l},z\right), \tag{17}
$$
where $L$ here denotes the batch size and can be adapted dynamically during training. An explicit example of this approach is presented in Section 3.3, in the context of batch augmentation to train a denoising autoencoder.
Weight-decay.
Schedules of regularization hyperparameters, e.g., the strength of the penalty on the $L2$ -norm of the weights, have also been empirically studied, for instance in the context of weight pruning [63]. The early work [64] investigated optimal regularization strategies through a variational approach akin to ours. More generally, hyperparameters of the regularization function $\tilde{g}$ can be directly included in the control vector $\bm{u}$ .
#### 2.4.2 Dynamic data selection
Accurately selecting training samples is a central challenge in modern machine learning. In heterogeneous datasets, e.g., composed of examples from multiple tasks or with varying levels of difficulty, the final performance of a model can be significantly influenced by the order in which samples are presented during training.
Task ordering.
The ability to learn new tasks without forgetting previously learned ones is crucial for both artificial and biological learners [65, 66]. Recent theoretical studies have assessed the relative effectiveness of various preâspecified task sequences [67, 68, 69, 70, 71]. In contrast, our framework allows to identify optimal task sequences in a variety of settings and was applied in [38] to derive interpretable taskâreplay strategies that minimize forgetting. The model in [67, 68, 38] is a special case of our formulation where each of the teacher vectors defines a different task $y_{m}=f^{*}_{\bm{w}^{*}_{m}}(\bm{x})$ , $m=1,\ldots,M$ , and $L=1$ . The student has $K=M$ hidden nodes and $H=M$ task-specific readout heads. When training on task $m$ , the loss function takes the simplified form
$$
\displaystyle\ell\left(\frac{{\bm{x}}^{\top}{\bm{w}_{*}}}{\sqrt{N}},\frac{{\bm
{x}}^{\top}{\bm{w}}}{\sqrt{N}},\frac{\bm{w}^{\top}\bm{w}}{N},{\bm{v}}\right)=
\hat{\ell}\left(\frac{{\bm{w}^{*}_{m}}\cdot{\bm{x}}}{\sqrt{N}},\frac{{\bm{w}}^
{\top}{\bm{x}}}{\sqrt{N}},\frac{\bm{w}^{\top}\bm{w}}{N},{\bm{v}}_{m}\right)\,. \tag{18}
$$
The task variable $m$ can then be treated as a control variable to identify optimal task orderings that minimize generalization error across tasks [38].
Curriculum learning.
When heterogeneous datasets involve a notion of relative sample difficulty, it is natural to ask whether training performance can be enhanced by using a curriculum, i.e., by presenting examples in a structured order based on their difficulty, rather than sampling them at random. This question has been theoretically explored in recent literature [29, 72, 73] and is investigated within our formulation in Section 3.1.
Data imbalance.
Many real-world datasets exhibit class imbalance, where certain classes are significantly over-represented [74]. Recent theoretical work has used statistical physics to study class-imbalance mitigation through under- and over-sampling in sequential data [75, 76]. Further aspects of data imbalance, such as relative representation imbalance and different sub-population variances, have been explored using a TS setting in [77, 78]. All these types of imbalance can be incorporated in our general formulation, e.g., by tilting the distribution of cluster memberships $p_{c}(\bm{c})$ , the cluster variances, and the alignment parameters $\bm{S}$ between teacher vectors and cluster centroids (see Eq. (10)). This framework would allow to investigate dynamical mitigation strategiesâsuch as optimal data ordering, adaptive loss reweighting, and learning-rate schedulesâaimed at restoring balance.
#### 2.4.3 Dynamic architectures
Dynamic architectures allow models to adjust their structure during training based on data or task demands, addressing some limitations of static models [79]. Several heuristic strategies have been proposed to dynamically adapt a networkâs architecture, e.g., to avoid overfitting or to facilitate knowledge transfer. Our framework enables the derivation of principled mechanisms for adapting the architecture during training across several settings.
Dropout.
Dropout is a widely adopted dynamic regularization technique in which random subsets of the network are deactivated during training to encourage robust, independent feature representations [80, 81]. While empirical studies have proposed adaptive dropout probabilities to enhance performance [82, 83], a theoretical understanding of optimal dropout schedules remains limited. In recent work, we introduced a twoâlayer network model incorporating dropout and analyzed the impact of fixed dropout rates [84]. As shown in Section 3.2, our general framework contains the model of [84] as a special case, enabling the derivation of principled dropout schedules.
Gating.
Gating functions modify the network architecture by selectively activating specific pathways, thereby modulating information flow and allocating computational resources based on input context. This principle improves model efficiency and expressiveness, and underlies diverse systems such as mixture of experts [85], squeeze-and-excitation networks [86], and gated recurrent units [87]. Gated linear networksâintroduced in [88] as context-gated models based on local learning rulesâhave been investigated in several theoretical works [89, 90, 91, 92]. Our framework offers the possibility to study dynamic gating and adaptive modulation, including gain and engagement modulation mechanisms [41], by controlling the hyperparameters of the gating functions. For instance, in teacher-student settings as in Eqs. (2) and (5), the model considered in [92] arises as a special case of our formulation, where $L=1$ and $f_{\bm{w},\bm{v}}(\bm{x})=\sum_{k=1}^{\lfloor K/2\rfloor}g_{k}(\bm{w}_{k}\cdot \bm{x})\,(\bm{w}_{\lfloor K/2\rfloor+k}\cdot\bm{x})$ with gating functions $g_{k}$ .
Dynamic attention.
Self-attention is the core building block of the transformer architecture [93]. Dynamic attention mechanisms enhance standard attention by adapting its structure in response to input properties or task requirements, for example, by selecting sparse token interactions [94], varying attention spans [95], or pruning attention heads dynamically [96, 97]. Recent theoretical works have introduced minimal models of dotâproduct attention that admit an analytic characterization [43, 98, 99]. These models can be incorporated into our framework to study adaptive attention dynamics. In particular, a multi-head single-layer dot-product attention model can be recovered by setting
$$
\displaystyle f_{\bm{w},\bm{v}}(\bm{x})=\sum_{h=1}^{H}v^{(h)}\bm{x}
\operatorname{softmax}\left(\frac{\bm{x}^{\top}\bm{w}^{(h)}_{\mathcal{Q}}{\bm{
w}^{(h)}_{\mathcal{K}}}^{\top}\bm{x}}{N}\right)\in\mathbb{R}^{N\times L}\;, \tag{19}
$$
where $\bm{w}^{(h)}_{\mathcal{Q}}\in\mathbb{R}^{N\times D_{H}}$ and $\bm{w}^{(h)}_{\mathcal{Q}}\in\mathbb{R}^{N\times D_{H}}$ denote the query and key matrices for the $h^{\rm th}$ head, with head dimension $D_{H}$ such that the total number of student vectors is $K=2HD_{H}$ . The value matrix is set to the identity, while the readout vector $\bm{v}\in\mathbb{R}^{H}$ acts as the output weights across heads. In teacher-student settings [98], the model in Eq. (19) is a special case of our formulation (see also [43]). Possible controls in this case include masking variables that dynamically prune attention heads, sparsify token interactions, or modulate context visibility, enabling adaptive structural changes to the model.
## 3 Applications
In this section, we present three different learning scenarios in which our framework allows to identify optimal learning strategies.
### 3.1 Curriculum learning
<details>
<summary>x1.png Details</summary>

### Visual Description
## Diagram: Teacher-Student Learning Model with Relevant and Irrelevant Features
### Overview
The image is a technical diagram illustrating a machine learning or statistical learning setup. It defines an input vector composed of relevant and irrelevant components and shows two models: a "Teacher" model that uses only the relevant features and a "Student" model that uses both relevant and irrelevant features. The diagram is structured into three distinct panels.
### Components/Axes
The diagram is divided into three rectangular panels with black borders.
1. **Left Panel (Input Definition):**
* **Title/Header:** `Input: x = (xâ, xâ) â âáŽșËŁÂČ`
* **Visual Elements:** A vertical column of circles representing input features.
* Top section: 7 green circles labeled vertically as **"Relevant"**.
* Bottom section: 7 red circles labeled vertically as **"Irrelevant"**.
* **Mathematical Definitions:**
* For the relevant component: `xâ ~ đ©(0, I_N)` with the annotation "Unit variance".
* For the irrelevant component: `xâ ~ đ©(0, âÎ I_N)`.
* A control parameter is defined: `u = Î` with the annotation "Control: variance" in blue text.
2. **Top-Right Panel (Teacher Model):**
* **Title/Header:** **"Teacher"**
* **Visual Elements:** A schematic of a single-layer neural network or linear classifier.
* Input: A column of 5 green circles (representing the relevant features `xâ`).
* Weights: Labeled `w*`.
* Output: A single circle.
* **Mathematical Formula:** `y = sign( (w* · xâ) / âN )`
3. **Bottom-Right Panel (Student Model):**
* **Title/Header:** **"Student"**
* **Visual Elements:** A schematic of a single-layer neural network or linear classifier.
* Input: A column of 5 green circles (relevant features `xâ`) and 5 red circles (irrelevant features `xâ`).
* Weights: Labeled `wâ` (connecting to green inputs) and `wâ` (connecting to red inputs).
* Output: A single circle.
* **Mathematical Formula:** `y = erf( (wâ · xâ + wâ · xâ) / (2âN) )`
### Detailed Analysis
* **Input Structure:** The input `x` is a matrix in `âáŽșËŁÂČ`, meaning it has `N` samples (or dimensions) and 2 feature groups. The two groups are statistically independent.
* **Relevant Features (`xâ`):** Drawn from a standard multivariate normal distribution with zero mean and identity covariance matrix `I_N`, implying unit variance and no correlation between features.
* **Irrelevant Features (`xâ`):** Drawn from a multivariate normal distribution with zero mean and a scaled covariance matrix `âÎ I_N`. The variance of each feature is controlled by the parameter `Î` (Delta), which is also defined as the control variable `u`.
* **Teacher Model:** This model represents the "true" or ideal process. It generates labels `y` using only the relevant features `xâ`. The output is the sign of the normalized dot product between a fixed weight vector `w*` and `xâ`. The normalization factor is `âN`.
* **Student Model:** This model attempts to learn the task. It has access to both relevant (`xâ`) and irrelevant (`xâ`) features, with separate weight vectors `wâ` and `wâ`. Its output uses the error function (`erf`) applied to a normalized sum of the weighted inputs. The normalization factor here is `2âN`.
### Key Observations
1. **Color Coding:** A consistent color scheme is used: green for relevant components (`xâ`, its weights `w*`/`wâ`, and its circles) and red for irrelevant components (`xâ`, its weights `wâ`, and its circles). Blue is used for the control parameter `u = Î`.
2. **Model Asymmetry:** The Teacher model is simpler, using a `sign` function and only relevant inputs. The Student model is more complex, using a smooth `erf` activation function and all inputs.
3. **Normalization:** Both models normalize their weighted sums by factors involving `âN`, but the Student's denominator (`2âN`) is twice as large as the Teacher's (`âN`).
4. **Spatial Layout:** The input definition is on the left, feeding conceptually into the two model panels on the right. The Teacher is positioned above the Student, suggesting a hierarchy or reference relationship.
### Interpretation
This diagram outlines a controlled experimental setup for studying learning in the presence of noise or irrelevant features. The core investigation appears to be: **How does the variance (Î) of irrelevant features affect a student model's ability to learn from relevant features when a teacher model provides labels based solely on those relevant features?**
* **The Teacher** represents the ground truth or optimal predictor, which is invariant to the irrelevant features `xâ`.
* **The Student** must learn the task but is "distracted" by the irrelevant features. The parameter `Î` controls the strength of this distraction. When `Î = 0`, the irrelevant features have zero variance (are constant), and the student's problem reduces to the teacher's. As `Î` increases, the irrelevant features become more variable and potentially more harmful to learning.
* The use of different activation functions (`sign` vs. `erf`) may be intentional to study the effect of model mismatch or the properties of different loss landscapes.
* This setup is characteristic of research in statistical learning theory, generalization, double descent, or the study of benign vs. harmful overfitting, where researchers analyze how model performance (e.g., generalization error) changes as a function of the noise level (`Î`) and the number of parameters or samples (`N`).
</details>
Figure 1: Illustration of the curriculum learning model studied in Section 3.1.
Curriculum learning (CL) refers to a variety of training protocols in which examples are presented in a curated orderâtypically organized by difficulty or complexity. In animal and human training, CL is widely used and extensively studied in behavioral research, demonstrating clear benefits [100, 101, 102]. For example, shaping âthe progressive introduction of subtasks to decompose a complex taskâis a common technique in animal training [6, 103]. By contrast, results on the efficacy of CL in machine learning remain sparse and less conclusive [104, 105]. Empirical studies across diverse settings have nonetheless demonstrated that curricula can outperform standard heuristic strategies [106, 107, 108].
Several theoretical studies have explored the benefits of curriculum learning in analytically tractable models. Easy-to-hard curricula have been shown to accelerate learning in convex settings [109, 110] and improve generalization in more complex nonconvex problems, such as XOR classification [111] or parity functions [112, 113]. However, these analyses typically focused on predefined heuristics, which may not be optimal. In particular, it remains unclear under what conditions an easyâtoâhard curriculum is truly optimal and what alternative strategies might outperform it when it is not. Moreover, although hyperparameter schedules have been shown to enhance curriculum learning empirically [49], a principled approach to their joint optimization remains largely unexplored.
Here, we focus on a prototypical model of curriculum learning introduced in [104] and recently studied analytically in [110], where high-dimensional learning curves for online SGD were derived. This model considers a binary classification problem in a TS setting where both teacher and student are perceptron (one-layer) networks. The input vectors consist of $L=2$ elementsârelevant directions $\bm{x}_{1}$ , which the teacher ( $M=1$ ) uses to generate labels $y=\operatorname{sign}({\bm{x}}_{1}\cdot{\bm{w}}_{*}/\sqrt{N})$ , and irrelevant directions $\bm{x}_{2}$ , which do not affect the labels For simplicity, we consider an equal proportion of relevant and irrelevant directions. It is possible to extend the analysis to arbitrary proportions as in [110].. The student network ( $K=2$ ) is given by
$$
f_{\bm{w}}(\bm{x})=\operatorname{erf}\left(\frac{{\bm{x}}_{1}\cdot{\bm{w}}_{1}
+{\bm{x}}_{2}\cdot{\bm{w}}_{2}}{2\sqrt{N}}\right)\,. \tag{20}
$$
As a result, the student does not know a priori which directions are relevant. The teacher vector is normalized such that $T_{11}=\bm{w}_{*}\cdot\bm{w}_{*}/N=2$ . All inputs are single-cluster zero-mean Gaussian variables and the sample difficulty is controlled by the variance $\Delta$ of the irrelevant directions, while the relevant directions are assumed to have unit variance (see Figure 1). We do not include label noise. We consider the squared loss $\ell=(y-f_{\bm{w}}(\bm{x}))^{2}/2$ and ridge regularization $\tilde{g}\left(\bm{w}^{\top}\bm{w}/N\right)=\lambda\left({\bm{w}}_{1}\cdot{\bm {w}}_{1}+{\bm{w}}_{2}\cdot{\bm{w}}_{2}\right)/(4N)$ , with tunable strength $\lambda\geq 0$ . An illustration of the model is presented in Figure 1. Full expressions for the ODEs governing the learning dynamics of the order parameters $M_{11}={\bm{w}}_{*}\cdot{\bm{w}}_{1}/N$ , $Q_{11}={\bm{w}}_{1}\cdot{\bm{w}}_{1}/N$ , $Q_{22}={\bm{w}}_{2}\cdot{\bm{w}}_{2}/N$ , and the generalization error are provided in Appendix A.1.
<details>
<summary>x2.png Details</summary>

### Visual Description
\n
## Multi-Panel Technical Figure: Curriculum Learning Analysis
### Overview
The image is a composite figure containing four subplots (labeled a, b, c, d) that analyze and compare three different training protocolsâCurriculum, Anti-Curriculum, and Optimalâacross various metrics over training time (α). The figure appears to be from a machine learning research paper investigating the effects of data presentation order on model performance and internal representations.
### Components/Axes
The figure is divided into four quadrants:
* **Top-Left (a):** A line chart plotting "Generalization error" (logarithmic y-axis) vs. "Training time α" (linear x-axis).
* **Top-Right (b):** A horizontal bar chart illustrating the "Difficulty protocol Î" over "Training time α".
* **Bottom-Left (c):** A line chart plotting "Cosine similarity with signal" (linear y-axis) vs. "Training time α", containing an inset zoom.
* **Bottom-Right (d):** A line chart plotting "Norm of irrelevant weights" (linear y-axis) vs. "Training time α".
**Common Legend (applies to plots a, c, d):**
* **Curriculum:** Blue dashed line with circle markers.
* **Anti-Curriculum:** Orange dashed line with square markers.
* **Optimal:** Black solid line with diamond markers.
**Axes Labels and Scales:**
* **X-axis (all plots):** "Training time α", ranging from 0 to 12.
* **Y-axis (Plot a):** "Generalization error", logarithmic scale. Major ticks at 2Ă10â»Âč, 3Ă10â»Âč, 4Ă10â»Âč.
* **Y-axis (Plot c):** "Cosine similarity with signal", linear scale from 0.0 to 1.0.
* **Y-axis (Plot d):** "Norm of irrelevant weights", linear scale from 1.0 to 4.0+.
* **Y-axis (Plot b):** "Difficulty protocol Î", categorical (Curriculum, Anti-Curriculum, Optimal).
### Detailed Analysis
**Plot a) Generalization Error vs. Training Time**
* **Trend Verification:** All three lines show a decreasing trend, indicating error reduction over training. The Curriculum line decreases most rapidly initially, then plateaus and is overtaken by the Optimal line. The Anti-Curriculum line decreases the slowest but continues a steady decline.
* **Data Points (Approximate):**
* **α=0:** All start at a high error (~5Ă10â»Âč).
* **α=2:** Curriculum ~2.5Ă10â»Âč, Optimal ~2.8Ă10â»Âč, Anti-Curriculum ~3.5Ă10â»Âč.
* **α=6:** Curriculum reaches a local minimum (~1.5Ă10â»Âč), then slightly increases. Optimal continues down to ~2.0Ă10â»Âč. Anti-Curriculum ~2.5Ă10â»Âč.
* **α=12:** Optimal achieves the lowest error (~1.2Ă10â»Âč). Curriculum and Anti-Curriculum converge near ~1.4Ă10â»Âč.
**Plot b) Difficulty Protocol Schematic**
* **Component Isolation:** This diagram defines the three protocols. The x-axis represents training time. Colors indicate task difficulty: **Cyan = Easy**, **Red = Hard**.
* **Curriculum:** Easy tasks first (cyan bar from α=0 to ~α=6), then Hard tasks (red bar from ~α=6 to α=12).
* **Anti-Curriculum:** Hard tasks first (red bar from α=0 to ~α=6), then Easy tasks (cyan bar from ~α=6 to α=12).
* **Optimal:** A mixed schedule. Starts with a short Easy phase (cyan, α=0 to ~α=2), followed by a long Hard phase (red, ~α=2 to ~α=8), and ends with another Easy phase (cyan, ~α=8 to α=12).
**Plot c) Cosine Similarity with Signal vs. Training Time**
* **Trend Verification:** All lines increase, indicating the model's learned representation becomes more aligned with the true signal over time. The inset shows a late-stage crossover.
* **Data Points (Approximate):**
* **α=0:** All start at 0.0.
* **α=2:** Curriculum ~0.7, Optimal ~0.65, Anti-Curriculum ~0.5.
* **α=6:** Curriculum ~0.9, Optimal ~0.85, Anti-Curriculum ~0.8.
* **α=12 (Main Plot):** Curriculum and Optimal converge near ~0.95. Anti-Curriculum is slightly lower, ~0.93.
* **Inset (α=8 to 12):** Provides a zoomed view. At α=8, Curriculum (~0.92) is highest. By α=12, Optimal (~0.95) has slightly surpassed Curriculum (~0.94), with Anti-Curriculum (~0.93) lowest.
**Plot d) Norm of Irrelevant Weights vs. Training Time**
* **Trend Verification:** This measures the magnitude of weights not contributing to the signal. Anti-Curriculum rises sharply and plateaus high. Optimal rises later to a medium plateau. Curriculum stays low the longest before a late, moderate rise.
* **Data Points (Approximate):**
* **α=0-2:** All norms are at baseline (~1.0).
* **Anti-Curriculum:** Begins rising sharply at α=1, plateaus near 4.2 by α=6.
* **Optimal:** Begins rising at α=2, plateaus near 2.8 by α=8.
* **Curriculum:** Remains near 1.0 until α=6, then rises to plateau near 2.4 by α=10.
### Key Observations
1. **Performance Crossover:** While the Curriculum protocol (easy-to-hard) leads to the fastest initial drop in generalization error (Plot a) and fastest gain in signal similarity (Plot c), the Optimal mixed schedule ultimately achieves the lowest final error.
2. **Protocol Impact on Internal Representation:** The Anti-Curriculum (hard-to-easy) protocol results in a significantly larger norm for irrelevant weights (Plot d), suggesting the model allocates more capacity to noise when faced with difficult tasks early in training.
3. **Late-Stage Dynamics:** The inset in Plot c reveals a subtle but important crossover where the Optimal protocol's representation quality surpasses the Curriculum's in the final stages of training.
4. **Schematic Clarity:** Plot b provides a clear visual definition of the experimental conditions, essential for interpreting the quantitative results in the other panels.
### Interpretation
This figure presents a Peircean investigation into the **causal relationship** between training data scheduling (the **sign** or protocol) and model learning outcomes (the **objects** of error, similarity, and weight norms). The data suggests that the order of task difficulty is not merely a heuristic but a critical factor shaping the learning trajectory.
* **Curriculum Learning (EasyâHard)** acts as a strong **index** for rapid initial learning, guiding the model efficiently toward a good solution. However, it may lead to premature convergence or suboptimal final representations, as hinted by the plateau in error and the late crossover in similarity.
* **Anti-Curriculum Learning (HardâEasy)** appears to be a **symbol** of inefficient learning. Starting with hard tasks forces the model to develop complex, potentially noisy representations early (high irrelevant weight norm), which then must be corrected, resulting in slower overall progress.
* **The Optimal Schedule** represents a **rheme** or hypothesis about an ideal balance. Its pattern (EasyâHardâEasy) suggests a beneficial structure: an initial easy phase for warm-up, a core hard phase for capacity building, and a final easy phase for refinement and consolidation. This schedule achieves the best final generalization, implying that strategic "scaffolding" and "consolidation" phases are key.
The anomaly is the Curriculum's early lead in similarity (Plot c) despite not having the lowest final error. This indicates that alignment with the signal alone is insufficient for optimal generalization; the management of irrelevant parameters (Plot d) is equally crucial. The figure collectively argues for principled, dynamically adjusted training protocols over simple monotonic difficulty schedules.
</details>
Figure 2: Learning dynamics for different difficulty schedules: curriculum (easy-to-hard), anti-curriculum (hard-to-easy) and the optimal one. a) Generalization error vs. training time $\alpha$ . b) Timeline of each schedule. c) Cosine similarity with the target signal $M_{11}/\sqrt{T_{11}Q_{11}}$ (inset zooms into the late-training regime). d) Squared norm of irrelevant weights $Q_{22}$ vs. $\alpha$ . Parameters: $\alpha_{F}=12$ , $\Delta_{1}=0$ , $\Delta_{2}=2$ , $\eta=3$ , $\lambda=0$ , $T_{11}=2$ . Initialization: $Q_{11}=Q_{22}=1$ , $M_{11}=0$ .
We consider a dataset composed of two difficulty levels: $50\$ âeasyâ examples ( $\Delta=\Delta_{1}$ ), and $50\$ âhardâ examples ( $\Delta=\Delta_{2}>\Delta_{1}$ ). We call curriculum the easy-to-hard schedule in which all easy samples are presented first, and anti-curriculum the opposite strategy (see Figure 2 b). We compute the optimal sampling strategy $\bm{u}(\alpha)=\Delta(\alpha)\in\{\Delta_{1},\Delta_{2}\}$ using Pontryaginâs maximum principle, as explained in Section 2.3.1. The constraint on the proportion of easy and hard examples in the training set is enforced via an additional Lagrange multiplier in the cost functional (Eq. (13)). As the final objective in Eq. (12) we use the misclassification error averaged over an equal proportion of easy and hard examples.
Good generalization requires balancing two competing objectives: maximizing the teacherâstudent alignment along relevant directionsâas measured by the cosine similarity with the signal $M_{11}/\sqrt{T_{11}Q_{11}}$ âand minimizing the norm of the studentâs weights along the irrelevant directions, $\sqrt{Q_{22}}$ . We observe that anti-curriculum favors the first objective, while curriculum the latter. This is shown in Figure 2, where we take constant learning rate $\eta=3$ and no regularization $\lambda=0$ . In this case, the optimal strategy is non-monotonic in difficulty, following an âeasy-hard-easyâ schedule, that balances the two objectives (see panels 2 c and 2 d), and achieves lower generalization error compared to the two monotonic strategies.
<details>
<summary>x3.png Details</summary>

### Visual Description
## Line Graphs: Regularization Impact and Optimal Learning Rate Schedule
### Overview
The image contains two distinct line charts, labeled **a)** and **b)**, presented side-by-side. Chart **a)** compares the final error of four different training strategies as a function of a regularization parameter. Chart **b)** plots the optimal learning rate over training time, with a visual indicator separating "Easy" and "Hard" phases of training.
### Components/Axes
**Chart a) - Left Panel**
* **Chart Type:** Line graph with markers.
* **Y-axis:** Label is "Final error". Scale is logarithmic, ranging from `10^-1` (0.1) to `2 x 10^-1` (0.2). Major ticks are at 0.1, 0.12, 0.14, 0.16, 0.18, and 0.2.
* **X-axis:** Label is "Regularization λ". Scale is linear, ranging from 0.00 to 0.30. Major ticks are at 0.00, 0.05, 0.10, 0.15, 0.20, 0.25, and 0.30.
* **Legend:** Located in the bottom-right quadrant of the chart area. It defines four data series:
1. `Curriculum`: Blue, dash-dot line with circle markers.
2. `Anti-Curriculum`: Orange, dashed line with square markers.
3. `Optimal (Î)`: Black, solid line with diamond markers.
4. `Optimal (Πand η)`: Green, solid line with 'x' markers.
**Chart b) - Right Panel**
* **Chart Type:** Line graph.
* **Y-axis:** Label is "Optimal learning rate η". Scale is linear, ranging from 1.0 to 5.0. Major ticks are at 1.0, 1.5, 2.0, 2.5, 3.0, 3.5, 4.0, 4.5, and 5.0.
* **X-axis:** Label is "Training time α". Scale is linear, ranging from 0 to 12. Major ticks are at 0, 2, 4, 6, 8, 10, and 12.
* **Phase Indicator:** A horizontal bar at the top of the chart area. The left segment (cyan) is labeled "Easy". The right segment (salmon/red) is labeled "Hard". The transition occurs at approximately `α = 6`.
### Detailed Analysis
**Chart a) - Regularization vs. Final Error**
* **Trend Verification & Data Points:**
* **Curriculum (Blue, circles):** Shows a clear, consistent upward trend. The line slopes upward from left to right.
* At λ=0.00, error â 0.155.
* At λ=0.15, error â 0.180.
* At λ=0.30, error â 0.205.
* **Anti-Curriculum (Orange, squares):** Shows a slight downward trend initially, then flattens.
* At λ=0.00, error â 0.162.
* At λ=0.15, error â 0.157.
* At λ=0.30, error â 0.160.
* **Optimal (Î) (Black, diamonds):** Shows a steady, shallow upward trend.
* At λ=0.00, error â 0.147.
* At λ=0.15, error â 0.153.
* At λ=0.30, error â 0.160.
* **Optimal (Πand η) (Green, 'x's):** Shows a shallow U-shaped curve, decreasing to a minimum before slightly increasing.
* At λ=0.00, error â 0.108.
* Minimum error occurs around λ=0.15, error â 0.102.
* At λ=0.30, error â 0.105.
**Chart b) - Optimal Learning Rate Schedule**
* **Trend Verification:** The line shows a rapid initial increase to a peak, followed by a steady decline, and then a sharp, discontinuous drop.
* **Data Points & Phases:**
* **Easy Phase (α â 0 to 6):** The optimal learning rate η starts around 4.2 at α=0, peaks at approximately η=4.8 near α=1, then declines steadily.
* **Transition:** At α=6, there is a sharp, vertical drop in the optimal learning rate from approximately η=2.4 to η=1.4.
* **Hard Phase (α â 6 to 12):** Following the drop, the learning rate continues a slow, linear decline from η=1.4 at α=6 to η=1.0 at α=12.
### Key Observations
1. **Performance Hierarchy (Chart a):** The "Optimal (Î and η)" strategy consistently achieves the lowest final error across all regularization values, followed by "Optimal (Î)". The "Curriculum" strategy performs the worst as regularization increases.
2. **Regularization Sensitivity (Chart a):** The "Curriculum" strategy is highly sensitive to regularization, with error increasing significantly as λ grows. The "Anti-Curriculum" and "Optimal (Î)" strategies are moderately sensitive. The "Optimal (Î and η)" strategy is the least sensitive, maintaining low error.
3. **Learning Rate Schedule (Chart b):** The optimal learning rate is not constant. It is high and dynamic during the "Easy" phase of training and undergoes a dramatic, step-wise reduction when transitioning to the "Hard" phase.
4. **Discontinuity (Chart b):** The sharp drop at α=6 is the most salient feature, indicating a fundamental shift in the optimal training dynamics at that point.
### Interpretation
The data suggests a sophisticated view of training dynamics in machine learning.
* **Chart a)** demonstrates that simply ordering data (Curriculum or Anti-Curriculum) is less effective than optimizing other hyperparameters (Î, and especially Î combined with η). The superior and robust performance of "Optimal (Î and η)" implies that jointly tuning the data ordering parameter (Î) and the learning rate (η) is crucial for achieving low error that is resilient to regularization strength. The poor performance of the standard Curriculum strategy with increasing regularization suggests it may lead to overfitting or poor generalization under those conditions.
* **Chart b)** provides a mechanistic insight. The "Easy" vs. "Hard" phase distinction, coupled with the learning rate schedule, suggests a two-stage training process. The initial high, peaking learning rate likely facilitates rapid exploration of the parameter space. The sharp drop at α=6 marks a transition pointâperhaps where the model has learned coarse features and must now fine-tune on harder examples or avoid overshooting minima. The subsequent low, decaying learning rate in the "Hard" phase is consistent with fine-grained convergence. This visualizes the concept that the optimal learning rate is not a fixed value but a schedule that should adapt to the training phase.
**Overall Synthesis:** Together, the charts argue for moving beyond simple curriculum learning. They advocate for an optimized, phase-aware training regimen where hyperparameters like the learning rate are dynamically adjusted in sync with the perceived difficulty of the training data, leading to more robust and effective models.
</details>
Figure 3: Simultaneous optimization of difficulty protocol $\Delta$ and learning rate $\eta$ in curriculum learning. a) Generalization error at the final time $\alpha_{F}=12$ , averaged over an equal fraction of easy and hard examples, as a function of the (rescaled) regularization $\bar{\lambda}=\lambda\eta$ for the three strategies presented in Figure 2, obtained optimizing over $\Delta$ at constant $\eta=3$ , and the optimal strategy (displayed in panel b for $\lambda=0$ ) obtained by jointly optimizing $\Delta$ and $\eta$ . Same parameters as Figure 2.
Furthermore, we observe that the optimal balance between these competing goals is determined by the interplay between the difficulty schedule and other problem hyperparameters such as regularization and learning rate. Figure 3 a shows the final generalization error as a function of the regularization strength (held constant during training) for curriculum (blue), anti-curriculum (orange), and the optimal schedule (black), at fixed learning rate. When the regularization is high ( $\lambda>0.2$ ), weight decay alone ensures norm suppression along the irrelevant directions, so the optimal strategy reduces to anti-curriculum.
We next explore how a timeâdependent learningârate schedule $\eta(\alpha)$ can be coupled with the curriculum to improve generalization. This corresponds to extending the control vector $\bm{u}(\alpha)=\left(\Delta(\alpha),\eta(\alpha)\right)$ , where both difficulty and learning rate schedules are optimized jointly. In Figure 3 a, we see that this joint optimization produces a substantial reduction in generalization error compared to any constantâ $\eta$ strategy. Interestingly, for all parameter settings considered, an easyâtoâhard curriculum becomes optimal once the learning rate is properly adjusted. Figure 3 b displays the optimal learning rate schedule $\eta(\alpha)$ at $\lambda=0$ : it begins with a warmâup phase, transitions to gradual annealing, and then undergoes a sharp drop precisely when the curriculum shifts from easy to hard samples. This behavior is intuitive, since learning harder examples benefits from a lower, more cautious learning rate. As demonstrated in Figure 10 (Appendix B), this combined schedule effectively balances both objectivesâmaximizing signal alignment and minimizing noise overfitting. These results align with the empirical learning rate scheduling employed in the numerical experiments of [111], where easier samples were trained with a higher (constant) learning rate and harder samples with a lower one. Importantly, our framework provides a principled derivation of the optimal joint schedule, thereby confirming and grounding prior empirical insights.
### 3.2 Dropout regularization
<details>
<summary>x4.png Details</summary>

### Visual Description
## Diagram: Teacher-Student Neural Network Architecture with Dropout and Rescaling
### Overview
The image is a technical diagram illustrating a teacher-student learning framework for neural networks, specifically depicting a process involving label noise, dropout during training, and a rescaling factor during testing. It consists of three distinct panels arranged horizontally, each showing a feedforward neural network architecture with associated mathematical notation and explanatory text.
### Components/Axes
The diagram is divided into three rectangular panels, each with a title and containing a neural network schematic.
**Panel 1 (Left): "Teacher"**
* **Network Structure:** A neural network with an input layer (labeled **x**), a single hidden layer with **M hidden nodes**, and a single output node.
* **Connections:** Green lines connect the input nodes to the hidden nodes. The connections from the hidden layer to the output node are labeled with the number **1**.
* **Weights:** The connections from input to hidden are labeled **w***.
* **Output:** The output of the hidden layer transformation is denoted as **Ï(x)**.
* **Equation:** Below the network: `y = Ï(x) + Ï_n z`
* **Label Noise Definition:** Text in red: "Label noise: z ~ N(0,1)" (indicating z is drawn from a standard normal distribution).
**Panel 2 (Center): "Student (at training step Ό)"**
* **Network Structure:** A neural network with an input layer (labeled **x**), a single hidden layer with **K hidden nodes**, and a single output node producing **Ć·**.
* **Connections:** Orange lines connect the input nodes to the hidden nodes. The connections from the hidden layer to the output node are labeled with the number **1**.
* **Weights:** The connections from input to hidden are labeled **w**.
* **Node-Activation Variables:** Purple squares are placed on the connections from the hidden layer to the output. They are labeled **r_Ό^(1)**, **r_Ό^(2)**, and **r_Ό^(K)**.
* **Text (Purple):** "Node-activation variables"
* **Equation:** `r_Ό^(1), r_Ό^(2), ..., r_Ό^(K) ~ Bernoulli(p_Ό)` (indicating each activation variable is an independent Bernoulli random variable with probability p_Ό).
**Panel 3 (Right): "Student (at testing time)"**
* **Network Structure:** Identical to the Student at training: input **x**, **K hidden nodes**, output **Ć·**.
* **Connections:** Orange lines connect input to hidden. The connections from the hidden layer to the output are colored **blue**.
* **Weights:** The connections from input to hidden are labeled **w**.
* **Rescaling Factor:** The blue output connections are collectively labeled **p_f**.
* **Text (Blue):** "Rescaling factor: p_f"
### Detailed Analysis
* **Teacher Network:** Represents a fixed, pre-trained model. It generates the target signal `Ï(x)` but the observed training label `y` is corrupted by additive Gaussian noise with standard deviation `Ï_n`.
* **Student Network (Training):** The student network learns from the noisy teacher. A key feature is the application of dropout to the hidden layer's output during training. Each of the K hidden nodes' contributions to the output is independently multiplied by a Bernoulli random variable `r_Ό^(k)` (which is 1 with probability `p_Ό` and 0 otherwise). This simulates randomly "dropping out" nodes.
* **Student Network (Testing):** At inference time, all K hidden nodes are active. To compensate for the fact that only a fraction `p_Ό` of nodes were active on average during training, the combined output from the hidden layer is scaled by a factor `p_f`. The diagram implies `p_f` is related to `p_Ό` (commonly `p_f = p_Ό` in standard dropout implementation).
### Key Observations
1. **Architectural Consistency:** The student network has the same structure (K hidden nodes) in both training and testing phases, but the processing of the hidden layer's output differs.
2. **Color Coding:** Colors are used functionally: Green for the teacher's weights, orange for the student's weights, purple for the stochastic training-time dropout variables, and blue for the deterministic testing-time rescaling factor.
3. **Parameter Notation:** The teacher uses `w*` (optimal/fixed weights), while the student uses `w` (learned weights). The teacher has `M` hidden nodes, while the student has `K` hidden nodes (where K may or may not equal M).
4. **Noise Model:** Label noise is explicitly modeled as additive Gaussian noise `N(0,1)` scaled by `Ï_n`.
### Interpretation
This diagram visually explains the mechanics of **knowledge distillation** combined with **dropout regularization**. The teacher network provides a "soft target" `Ï(x)`, which is a richer training signal than hard labels but is further corrupted by noise. The student network learns to mimic this signal.
The core insight is the depiction of dropout's train-test discrepancy. During training (center panel), the student's learning is stochastic due to the Bernoulli masking (`r_Ό`), which acts as a regularizer preventing co-adaptation of nodes. At test time (right panel), the network becomes deterministic, but its output must be rescaled by `p_f` to account for the fact that all nodes are now active, ensuring the expected magnitude of the output matches what was seen during training. The diagram effectively isolates and contrasts these two operational modes of the same student network. The presence of the teacher and label noise suggests this framework might be used for learning from noisy data or for model compression where a large teacher guides a smaller student.
</details>
Figure 4: Illustration of the dropout model studied in Section 3.2.
Dropout [80, 81] is a regularization technique designed to prevent harmful co-adaptations of hidden units, thereby reducing overfitting and enhancing the networkâs performance. During training, each node is independently kept active with probability $p$ and âdroppedâ (i.e., its output set to zero) otherwise, effectively sampling a random subnetwork at each iteration. At test time, the full network is used, which corresponds to averaging over the ensemble of all subnetworks and yields more robust predictions.
Dropout has become a cornerstone of modern neuralânetwork training [114]. While early works recommended keeping the activation probability fixedâtypically in the range $0.5$ - $0.8$ âthroughout training [80, 81], recent empirical studies propose varying this probability over time, using adaptive schedules to further enhance performance [115, 82, 83]. In particular, [82] showed that heuristic schedules that decrease the activation probability over time are analogous to easy-to-hard curricula and can lead to improved performance. Although adaptive dropout schedules have attracted practical interest, the conditions under which they outperform constant strategies remain poorly understood, and the theoretical foundations of their potential optimality are largely unexplored.
<details>
<summary>x5.png Details</summary>

### Visual Description
## Composite Figure: Training Dynamics of Dropout Methods
### Overview
The image is a composite figure containing four line charts arranged in a 2x2 grid, labeled a), b), c), and d). All charts share the same x-axis label, "Training time α", ranging from 0 to 5. The charts compare the performance and internal dynamics of different neural network training regimes, specifically focusing on dropout techniques. Subplots a), b), and c) compare three methods: "No dropout", "Constant (p=0.68)", and "Optimal". Subplot d) examines the "Activation probability p(α)" under different noise levels (Ïâ).
### Components/Axes
* **Common X-Axis (All subplots):** Label: "Training time α". Scale: Linear, from 0 to 5 with major ticks at 0, 1, 2, 3, 4, 5.
* **Subplot a):**
* **Y-Axis:** Label: "Generalization error". Scale: Logarithmic, with major ticks at 2Ă10â»ÂČ, 3Ă10â»ÂČ, 4Ă10â»ÂČ, 6Ă10â»ÂČ.
* **Legend (Top-right corner):**
* Orange squares, dashed line: "No dropout"
* Blue circles, dash-dot line: "Constant (p=0.68)"
* Black diamonds, solid line: "Optimal"
* **Subplot b):**
* **Y-Axis:** Label: "ÎÂČ" (Delta squared). Scale: Linear, from 0.0 to 0.8 with major ticks at 0.0, 0.2, 0.4, 0.6, 0.8.
* **Legend:** Not present. Line styles and colors are inferred to match subplot a).
* **Subplot c):**
* **Y-Axis:** Label: "Mââ/â(Qââ Tââ)". Scale: Linear, from 0.2 to 0.9 with major ticks at 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9.
* **Legend:** Not present. Line styles and colors are inferred to match subplot a).
* **Subplot d):**
* **Y-Axis:** Label: "Activation probability p(α)". Scale: Linear, from 0.4 to 1.0 with major ticks at 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0.
* **Legend (Bottom-left corner):**
* Teal squares, dotted line: "Ïâ = 0.1"
* Green circles, dashed line: "Ïâ = 0.2"
* Black diamonds, solid line: "Ïâ = 0.3"
* Pink crosses, solid line: "Ïâ = 0.5"
### Detailed Analysis
#### **Subplot a): Generalization Error vs. Training Time**
* **Trend Verification:** All three lines show a decreasing trend, indicating generalization error improves with training. The "No dropout" line (orange) decreases the slowest and plateaus at the highest error. The "Constant" line (blue) decreases faster. The "Optimal" line (black) decreases the fastest and reaches the lowest error.
* **Data Points (Approximate):**
* **α=0:** All lines start near 8Ă10â»ÂČ.
* **α=1:** No dropout â 5.5Ă10â»ÂČ; Constant â 4.5Ă10â»ÂČ; Optimal â 4.0Ă10â»ÂČ.
* **α=3:** No dropout â 3.5Ă10â»ÂČ; Constant â 2.5Ă10â»ÂČ; Optimal â 2.2Ă10â»ÂČ.
* **α=5:** No dropout â 3.2Ă10â»ÂČ; Constant â 1.8Ă10â»ÂČ; Optimal â 1.6Ă10â»ÂČ.
#### **Subplot b): ÎÂČ vs. Training Time**
* **Component Isolation (Lines matched to legend from a)):**
* **Orange (No dropout):** Starts highest (~0.95 at α=0.1), decreases steadily, plateaus around 0.25.
* **Black (Optimal):** Starts slightly lower than orange (~0.92 at α=0.1), decreases more steeply, ends around 0.08.
* **Blue (Constant):** Starts much lower (~0.63 at α=0.1), decreases rapidly, approaches near 0.0 by α=5.
* **Trend:** All lines show a decreasing, convex trend. The metric ÎÂČ is consistently ordered: No dropout > Optimal > Constant throughout training.
#### **Subplot c): Mââ/â(Qââ Tââ) vs. Training Time**
* **Component Isolation (Lines matched to legend from a)):**
* **Orange (No dropout) & Black (Optimal):** These two lines are nearly superimposed, especially after α=1. They start low (~0.27 at α=0.1), rise sharply, and plateau near 0.87.
* **Blue (Constant):** Follows a similar shape but is consistently lower. Starts at ~0.22, rises, and plateaus near 0.82.
* **Trend:** All lines show an increasing, concave trend that saturates. The "No dropout" and "Optimal" methods achieve a higher final value than the "Constant" method.
#### **Subplot d): Activation Probability p(α) vs. Training Time**
* **Trend Verification:** The behavior varies dramatically with Ïâ.
* **Ïâ = 0.1 (Teal, dotted):** Probability stays at 1.0 until αâ1.5, dips to a minimum of ~0.85 around α=3.5, then recovers back to 1.0 by α=4.5.
* **Ïâ = 0.2 (Green, dashed):** Probability stays at 1.0 until αâ1.5, then decreases monotonically, ending near 0.49.
* **Ïâ = 0.3 (Black, solid):** Probability starts decreasing earlier (αâ1.0), falls more steeply, ending near 0.44.
* **Ïâ = 0.5 (Pink, solid):** Probability begins decreasing almost immediately, falls the fastest, and ends at the lowest point (~0.42).
* **Key Observation:** Higher noise levels (Ïâ) cause the activation probability to drop earlier and more severely during training. The lowest noise level (Ïâ=0.1) shows a unique non-monotonic "dip and recovery" pattern.
### Key Observations
1. **Performance Hierarchy:** The "Optimal" dropout method consistently yields the lowest generalization error (a), followed by "Constant" dropout, with "No dropout" performing worst.
2. **Internal Metric Correlation:** The superior final performance of "Optimal" and "No dropout" in (a) correlates with their higher saturation value of the metric Mââ/â(Qââ Tââ) in (c). The "Constant" method has a lower value for this metric.
3. **Noise-Dependent Dynamics:** Subplot (d) reveals that the training dynamics of the activation probability are highly sensitive to the noise level Ïâ. There is a clear transition from a stable, high-probability regime (low Ïâ) to a rapidly decaying probability regime (high Ïâ).
### Interpretation
This figure provides a multi-faceted view of how different dropout strategies affect neural network training. The "Optimal" method, likely an adaptive or theoretically derived schedule, achieves the best generalization by balancing the trade-offs visualized in the other plots.
* **Subplot (a) is the primary outcome:** It shows the end-result benefit of the optimal strategy.
* **Subplots (b) and (c) offer mechanistic insights:** They track internal model quantities (ÎÂČ and a normalized overlap Mââ). The fact that "No dropout" and "Optimal" have similar, high values in (c) suggests they maintain a stronger signal or alignment in certain weight matrix components, which may be key to their generalization. The "Constant" dropout, while better than nothing, may overly regularize and suppress this signal.
* **Subplot (d) explains a potential mechanism for the "Constant" method's behavior:** If the constant dropout rate (p=0.68) corresponds to a specific effective noise level, its decaying activation probability could be a driver of the trends seen in (b) and (c). The non-monotonic curve for Ïâ=0.1 is particularly intriguing, suggesting a phase where the network initially relies on many features, prunes some during mid-training, and then re-engages them for fine-tuning.
**Overall, the data suggests that an "Optimal" dropout strategy is superior because it manages internal model dynamics (like feature activation and weight alignment) more effectively than a fixed dropout rate, leading to better final generalization.** The sensitivity to Ïâ highlights that the effectiveness of regularization is deeply tied to the scale of perturbation applied during training.
</details>
Figure 5: Learning dynamics with dropout regularization. a) Generalization error vs. training time $\alpha$ without dropout (orange), for constant activation probability $p=p_{f}=0.68$ (blue), and for the optimal dropout schedule with $p_{f}=0.678$ (black), at label noise $\sigma_{n}=0.3$ . b) Detrimental correlations between the studentâs hidden nodes, measured by $\tilde{\Delta}=(Q_{12}-M_{11}M_{21})/\sqrt{Q_{11}Q_{22}}$ , vs. $\alpha$ , at $\sigma_{n}=0.3$ . c) Teacher-student cosine similarity $M_{11}/\sqrt{Q_{11}T_{11}}$ vs. $\alpha$ , at $\sigma_{n}=0.3$ . d) Optimal dropout schedules for different label-noise levels. The black curve ( $\sigma_{n}=0.3$ ) shows the optimal schedule used in panels a - c. Parameters: $\alpha_{F}=5$ , $K=2$ , $M=1$ , $\eta=1$ . The teacher weights $\bm{w}^{*}$ are drawn i.i.d. from $\mathcal{N}(0,1)$ with $N=10000$ . The student weights are initialized to zero.
In [84], we introduced a prototypical model of dropout and derived analytic results for constant dropout probabilities. We showed that dropout reduces harmful node correlationsâquantified via order parametersâand consequently improves generalization. We further demonstrated that the optimal (constant) activation probability decreases as the variance of the label noise increases. In this section, we first recast the model of [84] within our general framework and then extend the analysis to optimal dropout schedules.
We consider a TS setup where both teacher and student networks are soft-committee machines [34], i.e., two-layer networks with untrained readout weights set to one. Specifically, the inputs $\bm{x}\in\mathbb{R}^{N}$ are taken to be standard Gaussian variables and the corresponding labels are produced via Eq. (2) with label noise variance $\sigma^{2}_{n}$ :
$$
\displaystyle y=f^{*}_{\bm{w}_{*}}(\bm{x})+\sigma_{n}\,z\;, \displaystyle z\sim\mathcal{N}(0,1)\;, \displaystyle f^{*}_{\bm{w}_{*}}(\bm{x})=\sum_{m=1}^{M}\operatorname{erf}\left
(\frac{\bm{w}_{*,m}\cdot{\bm{x}}}{\sqrt{N}}\right)\,. \tag{21}
$$
To describe dropout, at each training step $\mu$ we couple i.i.d. node-activation Bernoulli random variables $r^{(k)}_{\mu}\sim{\rm Ber}(p_{\mu})$ to each of the studentâs hidden nodes $k=1,\ldots,K$ :
$$
f^{\rm train}_{\bm{w}}(\bm{x}^{\mu})=\sum_{k=1}^{K}r^{(k)}_{\mu}\operatorname{
erf}\left(\frac{\bm{w}_{k}\cdot{\bm{x}}^{\mu}}{\sqrt{N}}\right)\,, \tag{22}
$$
so that node $k$ is active if $r^{(k)}_{\mu}=1$ . At testing time, the full network is used as
$$
f^{\rm test}_{\bm{w}}(\bm{x})=\sum_{k=1}^{K}p_{f}\operatorname{erf}\left(\frac
{\bm{w}_{k}\cdot{\bm{x}}}{\sqrt{N}}\right)\,. \tag{23}
$$
The rescaling factor $p_{f}$ ensures that the reduced activity during training is taken into account when testing. We consider the squared loss $\ell=(y-f_{\bm{w}}(\bm{x}))^{2}/2$ and no weight-decay regularization. The ODEs governing the order parameters $M_{km}$ and $Q_{jk}$ , as well as the resulting generalization error, are provided in Appendix A.2. These equations arise from averaging over the binary activation variables $r_{\mu}^{(k)}$ , so that the dropout schedule is determined by the timeâdependent activation probability $p(\alpha)$ .
For simplicity, we focus our analysis on the case $M=1$ and $K=2$ , although our considerations hold more generally. During training, assuming $T_{11}=1$ , each student weight vector can be decomposed as ${\bm{w}}_{i}=M_{i1}{\bm{w}}_{*,1}+\tilde{{\bm{w}}}_{i}$ , where $\tilde{\bm{w}}_{i}\perp\bm{w}_{*,1}$ denotes the uninformative component acquired due to noise in the inputs and labels. Generalization requires balancing two competing goals: improving the alignment of each hidden unit with the teacher, measured by $M_{i1}$ , and reducing correlations between their uninformative components, $\tilde{\bm{w}}_{1}$ and $\tilde{\bm{w}}_{2}$ , so that noise effects cancel rather than compound. We quantify these detrimental correlations by the observable $\tilde{\Delta}=(Q_{12}-M_{11}M_{21})/\sqrt{Q_{11}Q_{22}}$ . Figure 5 b compares a constantâdropout strategy ( $p=p_{f}=0.68$ , orange) with no dropout ( $p=p_{f}=1$ , blue) and shows that dropout sharply reduces $\tilde{\Delta}$ during training. Intuitively, without dropout, both nodes share identical noise realizations at each step, reinforcing their uninformative correlation; with dropout, nodes are from time to time trained individually, reducing correlations. Although dropout also slows the growth of the teacherâstudent cosine similarity (Figure 5 c) by reducing the number of updates per node, the large decrease in $\tilde{\Delta}$ leads to an overall lower generalization error (Figure 5 a).
To find the optimal dropout schedule, we treat the activation probability as the control variable, $u(\alpha)=p(\alpha)\in[0,1]$ . Additionally, we optimize over the final rescaling $p_{f}\in[0,1]$ to minimize the final error. We solve this optimalâcontrol problem using a direct multipleâshooting method implemented in CasADi (Section 2.3.2). Figure 5 shows the resulting optimal schedules for increasing labelânoise levels $\sigma_{n}$ . Each schedule exhibits an initial period with no dropout ( $p(\alpha)=1$ ) followed by a gradual decrease of $p(\alpha)$ . These strategies resemble those heuristically proposed in [82] but are obtained here via a principled procedure.
The order parameters of the theory suggest a simple interpretation of the optimal schedules. In the initial phase of training, it is beneficial to fully exploit the rapid increase in the teacher-student cosine similarity by keeping both nodes active (see Figure 5). Once the increase in cosine similarity plateaus, it becomes more advantageous to decrease the activation probability in order to mitigate negative correlations among the studentâs nodes. As a result, the optimal schedule achieves lower generalization error than any constantâdropout strategy.
Noisier tasks, corresponding to higher values of $\sigma_{n}$ , induce stronger detrimental correlations between the student nodes and therefore require a lower activation probability, as shown in [84] for the case of constant dropout. This observation remains valid for the optimal dropout schedules in Figure 5 d: as $\sigma_{n}$ grows, the initial noâdropout phase becomes shorter and the activation probability decreases more sharply. Conversely, at low label noise ( $\sigma_{n}=0.1$ ), the activation probability remains close to one and becomes non-monotonic in training time.
### 3.3 Denoising autoencoder
<details>
<summary>x6.png Details</summary>

### Visual Description
## Diagram: Two-layer Denoising Autoencoder (DAE) with Skip Connection
### Overview
The image is a technical diagram illustrating the architecture of a two-layer Denoising Autoencoder (DAE) that incorporates a skip connection. It visually decomposes the model's function into two primary components: a central "Bottleneck network" and a parallel "Skip connection." The diagram uses color-coding and mathematical notation to define the model's operation.
### Components/Axes
The diagram is organized into three distinct horizontal sections, each with a label at the top:
1. **Left Section (Black Label):** "Two-layer DAE"
* Contains the mathematical definition of the model's function: \( f_{\mathbf{w},b}(\tilde{\mathbf{x}}) = \)
2. **Center Section (Green Label):** "Bottleneck network"
* Depicts a neural network with three layers of nodes (circles).
* **Input Layer (Left):** Four black circles arranged vertically, with a vertical ellipsis (`âź`) between the second and third circles, indicating an arbitrary number of input nodes.
* **Hidden/Bottleneck Layer (Center):** Two green circles.
* **Output Layer (Right):** Four blue circles arranged vertically, with a vertical ellipsis (`âź`) between the second and third circles.
* **Connections:** Green lines connect all input nodes to all hidden nodes, and all hidden nodes to all output nodes.
* **Weight Labels:** The connections from input to hidden are labeled with a bold **`w`**. The connections from hidden to output are labeled with a bold **`w`** with a superscript T (**`wá”`**), indicating the transpose of the weight matrix.
* The output of this sub-network is labeled **`xÌ`** (x with a tilde).
3. **Right Section (Red Label):** "Skip connection"
* Depicts a vertical column of four black circles with a vertical ellipsis (`âź`) between the second and third circles, mirroring the input layer's structure.
* A red plus sign (`+`) and a red italicized **`b`** are positioned between the output of the bottleneck network and this column of nodes.
* The final output of the entire system, to the right of the skip connection nodes, is labeled **`xÌ`**.
### Detailed Analysis
The diagram explicitly defines the function \( f_{\mathbf{w},b}(\tilde{\mathbf{x}}) \) as the sum of two terms:
1. The output of the **Bottleneck network**, which processes the input **`xÌ`** through a hidden layer using weight matrices **`w`** and **`wá”`**.
2. A **Skip connection** term, which consists of a learnable bias vector **`b`** added to the original input **`xÌ`**.
**Spatial Grounding & Component Isolation:**
* The **legend/labels** are positioned directly above their corresponding components: "Two-layer DAE" (top-left), "Bottleneck network" (top-center), "Skip connection" (top-right).
* The **mathematical equation** flows from left to right, aligning with the visual components.
* The **color coding** is consistent: Green is used for the "Bottleneck network" label and its internal connections/hidden nodes. Red is used for the "Skip connection" label and the bias term **`b`**. Black is used for the primary function label, input/output nodes, and the final output symbol.
### Key Observations
* **Weight Tying:** The use of **`w`** for the encoder (input-to-hidden) and **`wá”`** (the transpose) for the decoder (hidden-to-output) indicates a "tied weights" architecture, a common technique in autoencoders to reduce parameters and enforce symmetry.
* **Skip Connection Nature:** The skip connection is not a direct identity pass-through of the input. Instead, it adds a bias term **`b`** to the input **`xÌ`**. This is a specific architectural choice, differing from a simple residual connection that would add the raw input **`xÌ`** itself.
* **Input/Output Notation:** The input and final output are both denoted as **`xÌ`**, suggesting the model's goal is to reconstruct its (possibly noisy) input.
### Interpretation
This diagram illustrates a specific variant of a denoising autoencoder designed for representation learning. The core function is performed by the **bottleneck network**, which compresses the input **`xÌ`** into a lower-dimensional hidden representation (the two green nodes) and then attempts to reconstruct it. The tied weights (**`w`** and **`wá”`**) impose a constraint that often leads to more robust features.
The **skip connection** adds a learned bias **`b`** to the original input before it is combined with the bottleneck network's output. This can help the model learn a residual correction or adjust the baseline of the reconstruction task. The overall architecture suggests a model that learns to extract essential features through compression while retaining a mechanism to directly modulate the input signal via the bias term. This could be useful for tasks where preserving certain global properties of the input is as important as learning its latent features.
</details>
Figure 6: Illustration of the denoising autoencoder model studied in Section 3.3.
<details>
<summary>x7.png Details</summary>

### Visual Description
\n
## Multi-Panel Line Chart: Training Dynamics Analysis
### Overview
The image is a composite figure containing four distinct line charts, labeled a), b), c), and d). Each chart plots a different performance or model parameter metric against a common x-axis variable, "Training time α". The charts collectively analyze the effects of a parameter ÎF and different optimization strategies (constant vs. optimal) on model training dynamics.
### Components/Axes
* **Common X-Axis (All Panels):** "Training time α". The scale runs from 0.0 to 0.8 with major tick marks at 0.1 intervals.
* **Panel a) Y-Axis:** "Optimal noise schedule". Scale from 0.0 to 0.8.
* **Panel b) Y-Axis:** "MSE improvement (%)". Scale from -40 to 30. A horizontal dashed line at 0% indicates the baseline.
* **Panel c) Y-Axis:** "Cosine similarity Ξ". Scale from 0.2 to 0.9.
* **Panel d) Y-Axis:** "Skip connection". Scale from 0.000 to 0.035.
* **Legends:**
* **Panel a) & b):** Located in the top-right corner. Contains six entries for different values of ÎF: 0.15 (yellow circle), 0.2 (blue star), 0.25 (green square), 0.3 (orange diamond), 0.35 (pink triangle up), 0.4 (grey triangle down).
* **Panel c):** Located in the bottom-right corner. Contains four entries: Ξâ,â^const (dark blue dashed line), Ξâ,â^const (light green dashed line), Ξâ,â^opt (dark blue solid line), Ξâ,â^opt (light green solid line).
* **Panel d):** Located in the top-left corner. Contains three entries: Target (black dotted line), Constant (green dash-dot line), Optimal (green solid line).
### Detailed Analysis
**Panel a) Optimal noise schedule vs. Training time α**
* **Trend:** For all ÎF values, the optimal noise schedule decreases as training time α increases. The rate of decrease is steeper for higher ÎF values.
* **Data Points (Approximate):**
* **ÎF = 0.15 (Yellow):** Starts near 0.0 at α=0.0, remains near 0.0 until αâ0.6, then rises slightly to ~0.05 at α=0.8.
* **ÎF = 0.2 (Blue):** Starts at ~0.26 at α=0.0, drops sharply to near 0.0 by α=0.15, and remains near 0.0.
* **ÎF = 0.25 (Green):** Starts at ~0.60 at α=0.0, decreases steadily, crossing 0.1 at αâ0.35, and approaches 0.0 by α=0.5.
* **ÎF = 0.3 (Orange):** Starts at ~0.77 at α=0.0, decreases, crossing 0.4 at αâ0.25 and 0.1 at αâ0.45.
* **ÎF = 0.35 (Pink):** Starts at ~0.78 at α=0.0, follows a similar but slightly higher path than ÎF=0.3.
* **ÎF = 0.4 (Grey):** Starts at ~0.78 at α=0.0, is the highest curve, crossing 0.4 at αâ0.4 and 0.1 at αâ0.55.
**Panel b) MSE improvement (%) vs. Training time α**
* **Trend:** The relationship is non-monotonic. For lower ÎF (0.15, 0.2), MSE improvement starts positive, peaks, then declines. For higher ÎF (0.25-0.4), improvement starts negative (worse than baseline), reaches a minimum, then rises sharply, often surpassing the lower ÎF curves at later training times (α > 0.6).
* **Data Points (Approximate):**
* **ÎF = 0.15 (Yellow):** Starts at ~7%, peaks at ~24% near α=0.45, declines to ~11% at α=0.8.
* **ÎF = 0.2 (Blue):** Starts at ~-4%, rises to a peak of ~29% near α=0.45, declines to ~15% at α=0.8.
* **ÎF = 0.25 (Green):** Starts at ~-23%, reaches a minimum of ~-25% near α=0.1, then rises, crossing 0% at αâ0.5 and reaching ~18% at α=0.8.
* **ÎF = 0.3 (Orange):** Starts at ~-35%, reaches a minimum of ~-42% near α=0.2, then rises steeply, crossing 0% at αâ0.57 and reaching ~19% at α=0.8.
* **ÎF = 0.35 (Pink):** Starts at ~-32%, minimum of ~-43% near α=0.25, crosses 0% at αâ0.62, reaches ~23% at α=0.8.
* **ÎF = 0.4 (Grey):** Starts at ~-27%, minimum of ~-40% near α=0.25, crosses 0% at αâ0.65, reaches ~28% at α=0.8.
**Panel c) Cosine similarity Ξ vs. Training time α**
* **Trend:** All four curves show a steady, sigmoidal increase in cosine similarity as training time α increases. The "optimal" (opt) strategies consistently achieve higher similarity than their "constant" (const) counterparts for the same Ξ index.
* **Data Points (Approximate):**
* **Ξâ,â^const (Dark Blue Dashed):** Starts at ~0.24 at α=0.0, rises to ~0.85 at α=0.8.
* **Ξâ,â^const (Light Green Dashed):** Starts at ~0.21 at α=0.0, rises to ~0.85 at α=0.8.
* **Ξâ,â^opt (Dark Blue Solid):** Starts at ~0.24 at α=0.0, rises more steeply, reaching ~0.92 at α=0.8.
* **Ξâ,â^opt (Light Green Solid):** Starts at ~0.21 at α=0.0, follows a path very close to Ξâ,â^opt, also reaching ~0.92 at α=0.8.
**Panel d) Skip connection vs. Training time α**
* **Trend:** The "Constant" and "Optimal" skip connection values increase linearly with training time α. The "Optimal" line has a steeper slope. The "Target" is a fixed horizontal line.
* **Data Points (Approximate):**
* **Target (Black Dotted):** Constant value of ~0.034 across all α.
* **Constant (Green Dash-Dot):** Starts at 0.0 at α=0.0, increases linearly to ~0.022 at α=0.8.
* **Optimal (Green Solid):** Starts at 0.0 at α=0.0, increases linearly with a steeper slope, reaching ~0.031 at α=0.8, approaching but not yet reaching the Target.
### Key Observations
1. **Trade-off in Panel b):** There is a clear trade-off between early-stage and late-stage performance. High ÎF values cause severe initial performance degradation (large negative MSE improvement) but lead to greater potential gains later in training.
2. **Convergence in Panel a):** Regardless of starting point, the optimal noise schedule for all ÎF values converges toward zero as training progresses (α > 0.6).
3. **Superiority of Optimal Strategies:** In both panels c) and d), the "opt" (optimal) strategies outperform their "const" (constant) counterparts, achieving higher cosine similarity and a faster approach to the target skip connection value.
4. **Crossover Point:** In panel b), a notable crossover occurs around α=0.65-0.7, where the initially poor-performing high-ÎF models begin to surpass the initially better low-ÎF models in MSE improvement.
### Interpretation
This figure likely analyzes training dynamics in a machine learning context, possibly for diffusion models or neural networks with skip connections. The parameter ÎF appears to control a noise or perturbation schedule.
* **Panels a) & b) together** suggest that a more aggressive initial noise schedule (high ÎF) is detrimental in the short term but beneficial for long-term model performance. The optimal strategy adapts this schedule, reducing noise as training stabilizes.
* **Panel c)** demonstrates that adaptive ("optimal") training strategies lead to better alignment (higher cosine similarity) between model components or representations compared to fixed ("constant") strategies.
* **Panel d)** shows that an adaptive approach for a skip connection parameter allows it to grow more efficiently toward a predefined target value, which is a common technique for stabilizing deep network training.
**Overall Narrative:** The data argues for the use of adaptive, time-varying training schedules (for noise and skip connections) over fixed ones. While aggressive adaptive schedules may incur an initial performance cost, they facilitate superior final model alignment and performance, as evidenced by higher late-stage MSE improvement and cosine similarity. The charts provide a quantitative basis for designing such schedules by showing the evolution of key metrics over training time.
</details>
Figure 7: a) Optimal noise schedule $\Delta$ vs. training time $\alpha$ . Each color marks a different value of the test noise level $\Delta_{F}$ . b) Percentage improvement in mean square error of the optimal strategy compared to the constant one at $\Delta(\alpha)=\Delta_{F}$ , computed as: $100(\operatorname{MSE}_{\rm const}(\alpha)-\operatorname{MSE}_{\rm opt}(\alpha ))/(\operatorname{MSE}_{\rm const}(0)-\operatorname{MSE}_{\rm const}(\alpha))$ . c) Cosine similarity $\theta_{k,k}=R_{k(1,k)}/\sqrt{Q_{kk}\Omega_{(1,k)(1,k)}}$ ( $k=1,2$ marked by different colors) vs. $\alpha$ for the optimal schedule (full lines) and the constant schedule (dashed lines), at $\Delta_{F}=0.25$ . d) Skip connection $b$ vs. $\alpha$ for the optimal schedule (full line) and the constant schedule (dashed line) at $\Delta_{F}=0.25$ . The dotted line marks the target value $b^{*}$ given by Eq. (26). Parameters: $K=C_{1}=2$ , $\alpha_{F}=0.8$ , $\eta=\eta_{b}=5$ , $\sigma=0.1$ , $N=1000$ , $g(z)=z$ . Initialization: $b=0$ . Other initial conditions are given in Eq. (92).
Denoising autoencoders (DAEs) are neural networks trained to reconstruct input data from their corrupted version, thereby learning robust feature representations [116, 117]. Recent developments in diffusion models have revived the interest in denoising tasks as a key component of the generative process [118, 119]. Several theoretical works have investigated the learning dynamics and generalization properties of DAEs. In the linear case, [120] showed that noise acts as a regularizer, biasing learning toward high-variance directions. Nonlinear DAEs were studied in [121], where exact asymptotics in high dimensions were derived. Relatedly, [122, 123] analyzed diffusion models parameterized by DAEs. [124] studied shallow reconstruction autoencoders in an online-learning setting closely related to ours.
A series of empirical works have considered noise schedules in the training of DAE. [125] showed that adaptive noise levels during training of DAEs promote learning multi-scale representations. Similarly, in diffusion models, networks are trained to denoise inputs at successive diffusion timestepsâeach linked to a specific noise level. Recent work [126] demonstrates that non-uniform sampling of diffusion time, effectively implementing a noise schedule, can further enhance performance. Additionally, data augmentation, where multiple independent corrupted samples are obtained for each clean input, is often employed [127]. However, identifying principled noise schedules and data augmentation strategies remains largely an open problem. In this section, we consider the prototypical DAE model studied in [121] and apply the optimal control framework introduced in Section 2 to find optimal noise and data augmentation schedules.
We consider input data $\bm{x}=(\bm{x}_{1},\bm{x}_{2})\in\mathbb{R}^{N\times 2}$ , where $\bm{x}_{1}\sim\mathcal{N}\left(\frac{\bm{\mu}_{1,c_{1}}}{\sqrt{N}},\sigma_{1,c _{1}}^{2}\bm{I}_{N}\right)$ , $c_{1}=1,\ldots,C_{1}$ , represents the clean input drawn from a Gaussian mixture of $C_{1}$ clusters, while $\bm{x}_{2}\sim\mathcal{N}(\bm{0},\bm{I}_{N})$ is additive standard Gaussian noise. We will take $\sigma_{1,c_{1}}=\sigma$ for all $c_{1}$ and equiprobable clusters unless otherwise stated. The network receives the noisy input $\tilde{\bm{x}}=\sqrt{1-\Delta}\,\bm{x}_{1}+\sqrt{\Delta}\,\bm{x}_{2}$ , where $\Delta>0$ controls the level of corruption. The denoising is performed via a two-layer autoencoder
$$
\displaystyle f_{\bm{w},b}(\tilde{\bm{x}})=\frac{\bm{w}}{\sqrt{N}}g\left(\frac
{\bm{w}^{\top}\tilde{\bm{x}}}{\sqrt{N}}\right)+b\,\tilde{\bm{x}}\;\in\mathbb{R
}^{N}\;, \tag{24}
$$
with tied weights $\bm{w}\in\mathbb{R}^{N\times K}$ , where $K$ is the dimension of the hidden layer, and a scalar trainable skip connection $b\in\mathbb{R}$ . The activation function $g$ is applied component-wise. The illustration in Figure 6 highlights the two components of the architecture: the bottleneck autoencoder network and the skip connection. In this unsupervised learning setting, the loss function is given by the squared reconstruction error between the clean input and the network output: $\mathcal{L}(\bm{w},b|\bm{x},\bm{c})=\|\bm{x}_{1}-f_{\bm{w},b}(\tilde{\bm{x}}) \|_{2}^{2}/2$ . This loss can be recast in the form of Eq. 4, as shown in [43]. The skip connection is trained via online SGD, i.e., $b^{\mu+1}=b^{\mu}-(\eta_{b}/N)\partial_{b}\mathcal{L}({\bm{w}}^{\mu},b^{\mu}|{ \bm{x}}^{\mu},{\bm{c}}^{\mu})$ .
We measure generalization via the mean squared error: $\operatorname{MSE}=\mathbb{E}_{\bm{x},\bm{c}}\left[\|\bm{x}-f_{\bm{w},b}( \tilde{\bm{x}})\|_{2}^{2}/2\right]$ . As shown in Appendix A.3, in the high-dimensional limit, the MSE is given by
$$
\displaystyle\begin{split}\text{MSE}=N\left[\sigma^{2}\left(1-b\sqrt{1-\Delta}
\right)^{2}+b^{2}\Delta\right]+\mathbb{E}_{\bm{x},\bm{c}}\left[\sum_{k,k^{
\prime}=1}^{K}Q_{kk^{\prime}}g(\tilde{\lambda}_{k})g(\tilde{\lambda}_{k^{
\prime}})-2\sum_{k=1}^{K}(\lambda_{1,k}-b\tilde{\lambda}_{k})g(\tilde{\lambda}
_{k})\right],\end{split} \tag{25}
$$
where we have defined the pre-activations $\tilde{\lambda}_{k}\equiv{\tilde{\bm{x}}}\cdot{\bm{w}}_{k}/\sqrt{N}$ and $\lambda_{1,k}={\bm{w}}_{k}\cdot{\bm{x}}_{1}/\sqrt{N}$ , and neglected a constant term. Note that the leading term in Eq. (25)âproportional to $N$ âis independent of the autoencoder weights $\bm{w}$ , and depends only on the skip connection $b$ and the noise level $\Delta$ . Therefore, the presence of the skip connection can improve the MSE by a contribution of order $\mathcal{O}_{N}(N)$ [122]. To leading order, the optimal skip connection that minimizes the MSE in Eq. (25) is given by
$$
b^{*}=\frac{\sqrt{(1-\Delta)}\,\sigma^{2}}{(1-\Delta)\,\sigma^{2}+\Delta}\;. \tag{26}
$$
The relevant order parameters in this model are $R_{k(1,c_{1})}$ and $Q_{kk^{\prime}}$ , where $k,k^{\prime}=1\ldots K$ and $c_{1}=1\ldots C_{1}$ (see Eq. (8) and (10)). In Appendix A.3, we provide closed-form expressions for the MSE and the ODEs describing the evolution of the order parameters.
<details>
<summary>x8.png Details</summary>

### Visual Description
## [Chart Type: Dual-Panel Line Chart with Inset Plot]
### Overview
The image displays two side-by-side line charts, labeled **a)** and **b)**, which appear to be from a technical or scientific paper. Both charts share the same x-axis variable, "Training time α". Chart **a)** plots "Optimal batch size" against α for different values of a parameter ÎF. Chart **b)** plots "MSE improvement (%)" against α for the same set of ÎF values and includes a smaller inset plot showing a relationship between a variable Î and an unnamed y-axis metric.
### Components/Axes
**Chart a) - Left Panel:**
* **Y-axis:** Label: "Optimal batch size". Scale: Linear, from 0 to 18, with major ticks every 2 units.
* **X-axis:** Label: "Training time α". Scale: Linear, from 0.0 to 1.2, with major ticks every 0.2 units.
* **Legend:** Located in the top-left corner. Contains six entries:
1. `ÎF = 0.1` (Yellow line, circle marker)
2. `ÎF = 0.3` (Blue line, 'x' marker)
3. `ÎF = 0.5` (Green line, square marker)
4. `ÎF = 0.7` (Orange line, diamond marker)
5. `ÎF = 0.9` (Pink line, triangle-up marker)
6. `Average batch size` (Black dashed line, no marker)
* **Reference Line:** A horizontal black dashed line at y=5, labeled "Average batch size".
**Chart b) - Right Panel:**
* **Y-axis:** Label: "MSE improvement (%)". Scale: Linear, from -20 to 20, with major ticks every 10 units. A horizontal dotted line is present at y=0.
* **X-axis:** Label: "Training time α". Scale: Linear, from 0.0 to 1.2, with major ticks every 0.2 units.
* **Legend:** The same color and marker scheme from chart a) is used here, though the legend itself is not repeated in this panel.
* **Inset Plot:** Located in the top-right corner of panel b).
* **Y-axis:** Unlabeled. Scale: Linear, from 0.0 to 10.0, with ticks every 2.5 units.
* **X-axis:** Label: "Î". Scale: Linear, from 0.0 to 0.8, with ticks every 0.2 units.
* **Data:** A single black line with circular markers, forming a peaked curve.
### Detailed Analysis
**Chart a) - Optimal Batch Size vs. Training Time:**
* **Trend Verification:** All five colored lines (for ÎF = 0.1 to 0.9) exhibit a clear upward trend as Training time α increases. The lines are step-like, suggesting discrete batch size values.
* **Data Points & Relationships:**
* At α = 0.0, the optimal batch size is low for all ÎF values, ranging from ~1 (for ÎF=0.5) to ~3 (for ÎF=0.9).
* As α increases, the optimal batch size for each ÎF series increases. The rate of increase is steeper for higher ÎF values.
* At the maximum shown α = 1.2:
* ÎF = 0.9 (Pink): ~10
* ÎF = 0.7 (Orange): ~14
* ÎF = 0.5 (Green): ~17
* ÎF = 0.3 (Blue): ~18
* ÎF = 0.1 (Yellow): ~15
* The "Average batch size" reference line is constant at 5. Most series cross above this line between α = 0.6 and α = 0.8.
**Chart b) - MSE Improvement vs. Training Time:**
* **Trend Verification:** The lines show varied initial values but generally trend upward as α increases, crossing from negative to positive MSE improvement.
* **Data Points & Relationships:**
* At α = 0.0:
* ÎF = 0.5 (Green): ~ -25% (lowest)
* ÎF = 0.3 (Blue): ~ -13%
* ÎF = 0.7 (Orange): ~ -13%
* ÎF = 0.9 (Pink): ~ -7%
* ÎF = 0.1 (Yellow): ~ -1% (closest to zero)
* All lines trend upward. They cross the 0% improvement line at different α values:
* ÎF = 0.1 (Yellow) crosses near α = 0.9.
* ÎF = 0.9 (Pink) crosses near α = 1.0.
* ÎF = 0.3 (Blue) and ÎF = 0.7 (Orange) cross near α = 1.05.
* ÎF = 0.5 (Green) crosses last, near α = 1.1.
* At α = 1.2, all series show positive improvement, ranging from ~1% (ÎF=0.9) to ~10% (ÎF=0.7).
* **Inset Plot Analysis:**
* The curve shows a clear peak. The y-value increases from ~0 at Î=0.0 to a maximum of ~10.0 at Î=0.6, then decreases back to ~1.5 at Î=0.8.
### Key Observations
1. **Inverse Relationship at Start:** In chart b), lower ÎF values (e.g., 0.1) start with MSE improvement near zero, while mid-range ÎF values (e.g., 0.5) start with the most negative improvement.
2. **Convergence at High α:** In chart a), the optimal batch sizes for different ÎF values diverge significantly as α increases. In chart b), the MSE improvement percentages for different ÎF values converge into a narrower positive range at high α.
3. **Peak Performance Indicator:** The inset plot in b) suggests there is an optimal value for the parameter Î (around 0.6) that maximizes the plotted metric.
4. **Step Function Behavior:** The optimal batch size in chart a) changes in discrete steps rather than a smooth curve, which may reflect practical constraints in batch size selection.
### Interpretation
This data explores the relationship between a training time parameter (α), a model or data parameter (ÎF), optimal batch size, and resulting model performance (MSE improvement).
* **Core Finding:** The "optimal" batch size is not static; it depends on both the training progress (α) and the underlying characteristic ÎF. As training progresses (α increases), larger batch sizes become optimal.
* **Performance Trade-off:** Chart b) reveals a critical insight: using a batch size that is not optimal for the current α and ÎF can lead to worse performance (negative MSE improvement) early in training. However, as training continues (α â 1.2), the system appears to recover, showing positive improvement regardless of the initial ÎF.
* **The Role of ÎF:** ÎF seems to control the sensitivity of the system. A mid-range ÎF (0.5) leads to the worst initial performance but also the steepest climb in optimal batch size. A very low ÎF (0.1) is stable initially but offers less potential for improvement.
* **Practical Implication:** The results argue for an **adaptive batch size strategy** during training. A fixed batch size (like the "Average" of 5) is suboptimal. The ideal strategy would start with a small batch size and increase it according to a schedule influenced by ÎF and α to maximize final performance and avoid early degradation. The inset plot likely helps in selecting or understanding the Î parameter that governs this adaptive process.
</details>
Figure 8: a) Optimal batch augmentation schedule vs. training time $\alpha$ for different values of the test noise level $\Delta=\Delta_{F}$ . All schedules have average batch size $\bar{B}=5$ . b) Percentage improvement of the optimal strategy compared to the constant one at $B(\alpha)=\bar{B}=5$ , computed as: $100(\operatorname{MSE}_{\rm const}(\alpha)-\operatorname{MSE}_{\rm opt}(\alpha ))/(\operatorname{MSE}_{\rm const}(0)-\operatorname{MSE}_{\rm const}(\alpha))$ . The inset shows the MSE improvement at the final time $\alpha_{F}=1.2$ as a function of $\Delta$ . Parameters: $K=C_{1}=2$ , $\eta=5$ , $\sigma=0.1$ , $g(z)=z$ . The skip connection $b$ is fixed ( $\eta_{b}=0$ ) to the optimal value in Eq. (26). Initial conditions are given in Eq. (92).
We start by considering the problem of finding the optimal denoising schedule $\Delta(\alpha)$ . Our goal is to minimize the final MSE, computed at the fixed test noise level $\Delta_{F}$ . To this end, we treat the noise level as the control variable $u(\alpha)=\Delta(\alpha)\in(0,1)$ , and we find the optimal schedule using a direct multiple-shooting method implemented in CasADi (Section 2.3.1). In the following analysis, we consider linear activation. Figure 7 a displays the optimal noise schedules for a range of test noise levels $\Delta_{F}$ . We observe that the optimal schedule typically features an initial decrease, followed by a moderate increase toward the end. At low $\Delta_{F}$ , the optimal schedule remains nearly flat and close to $\Delta=0$ before the final increase. Both the duration of the initial decreasing phase and the average noise level throughout the schedule increase with $\Delta_{F}$ . Figure 7 b shows that the optimal schedule improves the MSE by approximately $10$ - $30\$ over the constant schedule $\Delta(\alpha)=\Delta_{F}$ . The optimal denoising schedule achieves two key objectives. First, it enhances the reconstruction capability of the bottleneck network, leading to a higher cosine similarity between the hidden nodes of the autoencoder and the means of the Gaussian mixture defining the clean input distribution (panel 7 c). Second, it accelerates the convergence of the skip connection toward the target value $b^{*}$ in Eq. (26) (panel 7 d).
We then explore a setting that incorporates data augmentation, with inputs $\bm{x}=(\bm{x}_{1},\bm{x}_{2},\ldots,\bm{x}_{B+1})\in\mathbb{R}^{N\times B+1}$ , where $\bm{x}_{1}\sim\mathcal{N}\left(\frac{\bm{\mu}_{1,c_{1}}}{\sqrt{N}},\sigma^{2} \bm{I}_{N}\right)$ denotes the clean version of the input as before. We consider $B$ independent realizations of standard Gaussian noise $\bm{x}_{2},\ldots,\bm{x}_{B+1}\overset{\rm i.i.d.}{\sim}\mathcal{N}(\bm{0},\bm {I}_{N})$ . We can construct a batch of noisy inputs: $\tilde{\bm{x}}_{a}=\sqrt{1-\Delta}\,\bm{x}_{1}+\sqrt{\Delta}\,\bm{x}_{a+1}$ , $a=1,\ldots,B$ . The loss is averaged over the batch: $\mathcal{L}(\bm{w},b|\bm{x},\bm{c})=\sum_{a=1}^{B}\|\bm{x}_{1}-f_{\bm{w},b}( \tilde{\bm{x}}_{a})\|_{2}^{2}/(2B)$ . For simplicity, we take constant noise level $\Delta=\Delta_{F}$ and we fix the skip connection to its optimal value $b^{*}$ throughout training ( $\eta_{b}=0$ ). The ODEs can be extended to describe this setting, as shown in Appendix A.3.
We are interested in determining the optimal batch size schedule, that we take as our control variable $u(\alpha)=B(\alpha)\in\mathbb{N}$ . Specifically, we assume that we have access to a total budget of samples $B_{\rm tot}=\bar{B}\alpha_{F}N$ , where $\bar{B}$ is the average batch size available at each training time. We incorporate this constraint into the cost functional in Eq. (12) and solve the resulting optimization problem using CasADi. Figure 8 a shows the optimal batch size schedules varying the final noise level $\Delta_{F}$ . In all cases, the optimal schedule features a progressive increase in batch size throughout training, with only a moderate dependence on $\Delta_{F}$ . This corresponds to averaging the loss over a growing number of noise realizations, effectively reducing gradient variance and acting as a form of annealing that stabilizes learning in the later phases. This strategy leads to an MSE improvement of up to approximately $10\$ compared to the constant schedule preserving the total sample budget ( $B(\alpha)=\bar{B}$ ), as depicted in Figure 8 b. The inset shows that the final MSE gap is non-monotonic in $\Delta$ , with the highest improvement achieved at intermediate noise values.
<details>
<summary>x9.png Details</summary>

### Visual Description
## Line Charts and Image Comparison: Noise Schedule Optimization for Image Reconstruction
### Overview
The image is a composite figure containing three panels labeled a), b), and c). Panels a) and b) are line charts plotting metrics against training steps for different noise schedule parameters. Panel c) provides a visual comparison of image reconstruction results. The overall subject appears to be the optimization of a noise schedule (Î) during a training process to improve image reconstruction quality, measured by Mean Squared Error (MSE) improvement.
### Components/Axes
**Panel a):**
* **Chart Type:** Line chart.
* **X-axis:** Label: "Training step Ό". Scale: Linear, from 0 to 800, with major ticks every 100 steps.
* **Y-axis:** Label: "Noise schedule Î". Scale: Linear, from 0.0 to 0.5, with major ticks every 0.1.
* **Legend:** Located in the top-right corner. Contains four entries, each corresponding to a different value of ÎF:
* Yellow line with circle markers: ÎF = 0.1
* Blue line with 'x' markers: ÎF = 0.2
* Green line with square markers: ÎF = 0.3
* Orange line with diamond markers: ÎF = 0.4
**Panel b):**
* **Chart Type:** Line chart.
* **X-axis:** Label: "Training step Ό". Scale: Linear, from 0 to 800, with major ticks every 200 steps.
* **Y-axis:** Label: "MSE improvement (%)". Scale: Linear, from -30 to 40, with major ticks every 10%. A dashed horizontal line at 0% indicates the baseline.
* **Legend:** Implicitly matches the legend from panel a) by color and marker style. The four lines correspond to the same ÎF values (0.1, 0.2, 0.3, 0.4).
**Panel c):**
* **Content:** Two sets of four grayscale images each, arranged in a 2x4 grid.
* **Column Headers (Top Row):** "Original", "Corrupted", "Constant", "Optimal". These headers are repeated for the second set of images below.
* **Image Set 1 (Left):** Shows the reconstruction of a handwritten digit "0".
* **Image Set 2 (Right):** Shows the reconstruction of a handwritten digit "1".
### Detailed Analysis
**Panel a) - Noise Schedule Î vs. Training Step:**
* **Trend Verification:** All four lines follow a similar inverted-U shape. They start at a moderate value, rise to a peak between steps 200-300, and then decay towards zero by step 800.
* **Data Points (Approximate):**
* **ÎF = 0.1 (Yellow):** Starts ~0.23, peaks ~0.27 at step ~250, decays to ~0.01 by step 800.
* **ÎF = 0.2 (Blue):** Starts ~0.24, peaks ~0.34 at step ~250, decays to ~0.01 by step 800.
* **ÎF = 0.3 (Green):** Starts ~0.27, peaks ~0.45 at step ~250, decays to ~0.02 by step 800.
* **ÎF = 0.4 (Orange):** Starts ~0.29, peaks ~0.55 at step ~250, decays to ~0.02 by step 800.
* **Key Observation:** Higher ÎF values result in a higher peak noise schedule Î. The peak occurs at approximately the same training step (~250) for all series.
**Panel b) - MSE Improvement (%) vs. Training Step:**
* **Trend Verification:** All lines show a similar pattern: initial fluctuation near 0%, a significant dip into negative improvement (worsening) between steps 100-400, followed by a recovery and rise into positive improvement after step 400.
* **Data Points (Approximate):**
* **ÎF = 0.1 (Yellow):** Shows the smallest dip (~ -10% at step ~300) and the smallest final improvement (~ +3% at step 800).
* **ÎF = 0.2 (Blue):** Dips to ~ -12% at step ~300, recovers to ~ +9% at step 800.
* **ÎF = 0.3 (Green):** Dips to ~ -25% at step ~350, recovers strongly to ~ +18% at step 800.
* **ÎF = 0.4 (Orange):** Shows the most extreme behavior. Dips deepest to ~ -30% at step ~350, then recovers most strongly, reaching ~ +42% at step 800.
* **Key Observation:** There is a clear trade-off. Higher ÎF values cause a more severe temporary degradation in performance (deeper negative dip) but lead to a much greater final improvement in MSE. The crossover point to net positive improvement occurs later for higher ÎF (around step 450 for ÎF=0.4 vs. step 400 for ÎF=0.1).
**Panel c) - Visual Reconstruction Comparison:**
* **Content Details:**
* **Original:** Clean, high-contrast images of digits "0" and "1".
* **Corrupted:** The same digits heavily obscured by what appears to be dense, random noise. The digit structure is barely discernible.
* **Constant:** Reconstructions using a "Constant" noise schedule (presumably ÎF=0 or a fixed value). The digits are recognizable but blurry and retain significant noise artifacts.
* **Optimal:** Reconstructions using the optimized noise schedule (likely corresponding to the best-performing ÎF from the charts, e.g., ÎF=0.4). These images are noticeably sharper and clearer than the "Constant" versions, with better-defined edges and less background noise, closely resembling the "Original".
### Key Observations
1. **Parameter Sensitivity:** The system's performance is highly sensitive to the ÎF parameter. ÎF=0.4 yields the best final MSE improvement but at the cost of the worst mid-training performance.
2. **Training Phase Duality:** The training process exhibits two distinct phases: an initial "destructive" phase where performance worsens (negative MSE improvement), followed by a "constructive" phase where performance improves significantly.
3. **Visual-Quantitative Correlation:** The "Optimal" images in panel c) visually confirm the quantitative improvement shown in panel b). The superior sharpness of the "Optimal" reconstructions aligns with the high positive MSE improvement percentages for higher ÎF values.
### Interpretation
The data demonstrates the effectiveness of a dynamically scheduled noise parameter (Î) in an iterative image reconstruction or denoising task. The core insight is that **introducing more noise (higher ÎF) during the critical mid-training phase (around steps 200-400), while detrimental in the short term, ultimately guides the model to a better solution.**
* **Peircean Investigation:** The charts suggest a causal relationship: the optimized noise schedule (panel a) is the cause, and the improved MSE (panel b) and visual quality (panel c) are the effects. The "Constant" schedule acts as a control, proving that the dynamic schedule is the key variable.
* **Underlying Mechanism:** This pattern is characteristic of optimization processes that escape local minima. The initial performance dip may represent the model being pushed out of a suboptimal solution (a local minimum) by the increased noise. The subsequent recovery and strong improvement indicate the model finding a deeper, more robust minimum in the loss landscape.
* **Practical Implication:** For practitioners, this indicates a trade-off between training stability and final performance. Choosing a higher ÎF requires patience through a period of apparent degradation but promises superior results. The optimal ÎF value (0.4 in this case) maximizes the final outcome, suggesting the system benefits from a more aggressive exploration of the solution space early in training.
</details>
Figure 9: a) Optimal noise schedule $\Delta$ as a function of the training step $\mu$ for the MNIST dataset with only $0 0$ s and $1$ s. b) Percentage improvement in test mean square error of the optimal strategy compared to the constant one at $\Delta=\Delta_{F}$ . Each curve is averaged over $10$ random realizations of the training set. c) Examples of images for $\Delta_{F}=0.4$ : original, corrupted, denoised with the constant schedule $\Delta=\Delta_{F}$ , and denoised with the optimal schedule. Parameters: $K=C_{1}=2$ , $\alpha_{F}=1$ , $\eta=\eta_{b}=5$ , $\sigma=0.1$ , $N=784$ , $g(z)=z$ . Initialization: $b=0$ . Other initial conditions and parameters are given in Eq. (92).
We now demonstrate the applicability of our framework to real-world data by focusing on the MNIST dataset, which consists of labeled $28\times 28$ grayscale images of handwritten digits from $0 0$ to $9$ . For simplicity, we restrict our analysis to the digits $0 0$ and $1$ . To apply our framework, we numerically estimate the mean vectors ${\bm{\mu}}_{1,1}$ and ${\bm{\mu}}_{1,2}$ , corresponding to the digit classes $0 0$ and $1$ , respectively, as well as the standard deviations $\sigma_{1,1}$ and $\sigma_{1,2}$ . For additional details and initial conditions, see Appendix B. While our method could be extended to include the full covariance matrices, this would result in more involved dynamical equations [121, 128], which we leave for future work.
Considering learning trajectories with $\alpha_{F}=1$ , we use our theoretical framework to identify the optimal noise schedule $\Delta$ for different values of the testing noise $\Delta_{F}$ . The resulting schedules are shown in Fig. 9, and all exhibit a characteristic pattern: an initial increase in noise followed by a gradual decrease toward the end of the training trajectory. As expected, higher values of the testing noise $\Delta_{F}$ lead to overall higher noise levels throughout the schedule.
We then use these schedules to train a DAE with $K=2$ on a randomly selected training set of $P=784$ images (corresponding to $\alpha_{F}=P/N=1$ ). In Fig. 9 b, we compare the test-set MSE percent improvement relative to the constant strategy $\Delta=\Delta_{F}$ . We observe that the optimal noise schedule yields improvements of up to approximately $40\$ . This improvement is also apparent in the denoised images shown in Fig. 9 c. These results highlight the practical benefits of optimizing the noise schedule, confirming the applicability of our theoretical framework to real data.
## 4 Discussion
We have introduced a general framework for optimal learning that combines statistical physics with control theory to identify optimal training protocols. We have formulated the design of learning schedules as an OC problem on the low-dimensional dynamics of order parameters in a general two-layer neural network model trained with online SGD that captures a broad range of learning scenarios. The applicability of this framework was illustrated through several examples spanning hyperparameter tuning, architecture design, and data selection. We have then thoroughly investigated optimal training protocols in three representative settings: curriculum learning, dropout regularization, and denoising autoencoders.
We have consistently found that optimal training protocols outperform standard heuristics and can exhibit highly nontrivial structures that would be difficult to guess a priori. In curriculum learning, we have shown that non-monotonic difficulty schedules can outperform both easy-to-hard and hard-to-easy curricula. In dropout-regularized networks, the optimal schedule delayed the onset of regularization, exploiting the early phase to increase signal alignment before suppressing harmful co-adaptations. Optimal noise schedules for denoising autoencoders enhanced the reconstruction ability of the network while speeding up the training of the skip connection.
Interestingly, the dynamics of the order parameters often revealed interpretable structures in the resulting protocols a posteriori. Indeed, the order parameters allow to identify fundamental learning trade-offsâfor instance, alignment with informative directions versus suppression of noise fittingâwhich determine the structure of the optimal protocols. Our framework further enables the joint optimization of multiple controls, revealing synergies between meta-parameters, for example, how learning rate modulation can compensate for shifts in task difficulty.
Our framework can be extended in several directions. As detailed in Section 2.4, the current formulation already accommodates a variety of learning settings beyond those investigated here, including dynamic architectural features such as gating and attention. A first natural extension would involve considering more realistic data models [129, 18, 130, 22] to investigate how data structure affects optimal schedules. It would also be relevant to extend the OC framework introduced here to batch learning settings allowing to study how training schedules affect the interplay between memorization and generalization, e.g., via dynamical mean-field theory [25, 26, 131]. Additionally, it would be relevant to extend the analysis to deep and overparametrized architectures [28, 132]. Finally, the discussion in Section 3.3 on optimal noise schedules could be extended to generative settings such as diffusion models, enabling the derivation of optimal noise injection protocols [133]. Such connection could be explored within recently proposed minimal models of diffusion-based generative models [123].
Our framework can also be applied to optimize alternative training objectives. While we focused here on minimizing the final generalization error, other criteriaâsuch as fairness metrics in imbalanced datasets, robustness under distribution shift, or computational efficiencyâcan be incorporated within the same formalism. Finally, while we considered gradient-based learning rules, it would be interesting to explore biologically plausible update mechanisms or constraints on control signals inspired by cognitive or neural resource limitations [134, 135, 136].
### Acknowledgments
We thank Stefano Sarao Mannelli and Antonio Sclocchi for helpful discussions. We are grateful to Hugo Cui for useful feedback on the manuscript. This work was supported by a Leverhulme Trust International Professorship grant (Award Number: LIP-2020-014) and by the Simons Foundation (Award Number: 1141576).
## References
- [1] Yoshua Bengio. Practical recommendations for gradient-based training of deep architectures. In Neural networks: Tricks of the trade: Second edition, pages 437â478. Springer, 2012.
- [2] Amitai Shenhav, Matthew M Botvinick, and Jonathan D Cohen. The expected value of control: an integrative theory of anterior cingulate cortex function. Neuron, 79(2):217â240, 2013.
- [3] Matthew M. Botvinick and Jonathan D. Cohen. The computational and neural basis of cognitive control: Charted territory and new frontiers. Cognitive Science, 38(6):1249â1285, 2014.
- [4] Sebastian Musslick and Jonathan D Cohen. Rationalizing constraints on the capacity for cognitive control. Trends in cognitive sciences, 25(9):757â775, 2021.
- [5] Brett D Roads, Buyun Xu, June K Robinson, and James W Tanaka. The easy-to-hard training advantage with real-world medical images. Cognitive Research: Principles and Implications, 3:1â13, 2018.
- [6] Burrhus Frederic Skinner. The behavior of organisms: An experimental analysis. BF Skinner Foundation, 2019.
- [7] Luca Franceschi, Paolo Frasconi, Saverio Salzo, Riccardo Grazzi, and Massimiliano Pontil. Bilevel programming for hyperparameter optimization and meta-learning. In International conference on machine learning, pages 1568â1577. PMLR, 2018.
- [8] Frank Hutter, Lars Kotthoff, and Joaquin Vanschoren. Automated machine learning: methods, systems, challenges. Springer Nature, 2019.
- [9] James Bergstra and Yoshua Bengio. Random search for hyper-parameter optimization. The journal of machine learning research, 13(1):281â305, 2012.
- [10] Jasper Snoek, Hugo Larochelle, and Ryan P Adams. Practical bayesian optimization of machine learning algorithms. Advances in neural information processing systems, 25, 2012.
- [11] Dougal Maclaurin, David Duvenaud, and Ryan Adams. Gradient-based hyperparameter optimization through reversible learning. In International conference on machine learning, pages 2113â2122. PMLR, 2015.
- [12] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126â1135. PMLR, 2017.
- [13] Andreas Engel. Statistical mechanics of learning. Cambridge University Press, 2001.
- [14] Yasaman Bahri, Jonathan Kadmon, Jeffrey Pennington, Sam S Schoenholz, Jascha Sohl-Dickstein, and Surya Ganguli. Statistical mechanics of deep learning. Annual review of condensed matter physics, 11(1):501â528, 2020.
- [15] Florent Krzakala and Lenka ZdeborovĂĄ. Les houches 2022 special issue. Journal of Statistical Mechanics: Theory and Experiment, 2024(10):101001, 2024.
- [16] Jean Barbier, Florent Krzakala, Nicolas Macris, LĂ©o Miolane, and Lenka ZdeborovĂĄ. Optimal errors and phase transitions in high-dimensional generalized linear models. Proceedings of the National Academy of Sciences, 116(12):5451â5460, 2019.
- [17] Hugo Cui, Florent Krzakala, and Lenka Zdeborova. Bayes-optimal learning of deep random networks of extensive-width. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 6468â6521. PMLR, 23â29 Jul 2023.
- [18] Bruno Loureiro, Cedric Gerbelot, Hugo Cui, Sebastian Goldt, Florent Krzakala, Marc Mezard, and Lenka ZdeborovĂĄ. Learning curves of generic features maps for realistic datasets with a teacher-student model. Advances in Neural Information Processing Systems, 34:18137â18151, 2021.
- [19] Francesca Mignacco, Florent Krzakala, Yue Lu, Pierfrancesco Urbani, and Lenka Zdeborova. The role of regularization in classification of high-dimensional noisy Gaussian mixture. In Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 6874â6883. PMLR, 13â18 Jul 2020.
- [20] Dominik Schröder, Daniil Dmitriev, Hugo Cui, and Bruno Loureiro. Asymptotics of learning with deep structured (random) features. In Forty-first International Conference on Machine Learning, 2024.
- [21] Federica Gerace, Bruno Loureiro, Florent Krzakala, Marc Mezard, and Lenka Zdeborova. Generalisation error in learning with random features and the hidden manifold model. In Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 3452â3462. PMLR, 13â18 Jul 2020.
- [22] Urte Adomaityte, Gabriele Sicuro, and Pierpaolo Vivo. Classification of superstatistical features in high dimensions. In 2023 Conference on Neural Information Procecessing Systems, 2023.
- [23] Qianyi Li and Haim Sompolinsky. Statistical mechanics of deep linear neural networks: The backpropagating kernel renormalization. Phys. Rev. X, 11:031059, Sep 2021.
- [24] Sebastian Goldt, Madhu Advani, Andrew M Saxe, Florent Krzakala, and Lenka ZdeborovĂĄ. Dynamics of stochastic gradient descent for two-layer neural networks in the teacher-student setup. Advances in neural information processing systems, 32, 2019.
- [25] Francesca Mignacco, Florent Krzakala, Pierfrancesco Urbani, and Lenka ZdeborovĂĄ. Dynamical mean-field theory for stochastic gradient descent in gaussian mixture classification. Advances in Neural Information Processing Systems, 33:9540â9550, 2020.
- [26] Cedric Gerbelot, Emanuele Troiani, Francesca Mignacco, Florent Krzakala, and Lenka Zdeborova. Rigorous dynamical mean-field theory for stochastic gradient descent methods. SIAM Journal on Mathematics of Data Science, 6(2):400â427, 2024.
- [27] Yehonatan Avidan, Qianyi Li, and Haim Sompolinsky. Unified theoretical framework for wide neural network learning dynamics. Phys. Rev. E, 111:045310, Apr 2025.
- [28] Blake Bordelon and Cengiz Pehlevan. Self-consistent dynamical field theory of kernel evolution in wide neural networks. Advances in Neural Information Processing Systems, 35:32240â32256, 2022.
- [29] Luca Saglietti, Stefano Mannelli, and Andrew Saxe. An analytical theory of curriculum learning in teacher-student networks. In Advances in Neural Information Processing Systems, volume 35, pages 21113â21127. Curran Associates, Inc., 2022.
- [30] Jin Hwa Lee, Stefano Sarao Mannelli, and Andrew M Saxe. Why do animals need shaping? a theory of task composition and curriculum learning. In International Conference on Machine Learning, pages 26837â26855. PMLR, 2024.
- [31] Younes Strittmatter, Stefano S Mannelli, Miguel Ruiz-Garcia, Sebastian Musslick, and Markus Spitzer. Curriculum learning in humans and neural networks, Mar 2025.
- [32] Michael Biehl and Holm Schwarze. Learning by on-line gradient descent. Journal of Physics A: Mathematical and general, 28(3):643, 1995.
- [33] David Saad and Sara A Solla. Exact solution for on-line learning in multilayer neural networks. Physical Review Letters, 74(21):4337, 1995.
- [34] David Saad and Sara A Solla. On-line learning in soft committee machines. Physical Review E, 52(4):4225, 1995.
- [35] Megan C Engel, Jamie A Smith, and Michael P Brenner. Optimal control of nonequilibrium systems through automatic differentiation. Physical Review X, 13(4):041032, 2023.
- [36] Steven Blaber and David A Sivak. Optimal control in stochastic thermodynamics. Journal of Physics Communications, 7(3):033001, 2023.
- [37] Luke K Davis, Karel Proesmans, and Ătienne Fodor. Active matter under control: Insights from response theory. Physical Review X, 14(1):011012, 2024.
- [38] Francesco Mori, Stefano Sarao Mannelli, and Francesca Mignacco. Optimal protocols for continual learning via statistical physics and control theory. In International Conference on Learning Representations (ICLR), 2025.
- [39] David Saad and Magnus Rattray. Globally optimal parameters for on-line learning in multilayer neural networks. Physical review letters, 79(13):2578, 1997.
- [40] Magnus Rattray and David Saad. Analysis of on-line training with optimal learning rates. Physical Review E, 58(5):6379, 1998.
- [41] Rodrigo Carrasco-Davis, Javier MasĂs, and Andrew M Saxe. Meta-learning strategies through value maximization in neural networks. arXiv preprint arXiv:2310.19919, 2023.
- [42] Yujun Li, Rodrigo Carrasco-Davis, Younes Strittmatter, Stefano Sarao Mannelli, and Sebastian Musslick. A meta-learning framework for rationalizing cognitive fatigue in neural systems. In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 46, 2024.
- [43] Hugo Cui. High-dimensional learning of narrow neural networks. Journal of Statistical Mechanics: Theory and Experiment, 2025(2):023402, 2025.
- [44] Elizabeth Gardner and Bernard Derrida. Three unfinished works on the optimal storage capacity of networks. Journal of Physics A: Mathematical and General, 22(12):1983, 1989.
- [45] H. S. Seung, H. Sompolinsky, and N. Tishby. Statistical mechanics of learning from examples. Phys. Rev. A, 45:6056â6091, Apr 1992.
- [46] Maria Refinetti, StĂ©phane DâAscoli, Ruben Ohana, and Sebastian Goldt. Align, then memorise: the dynamics of learning with feedback alignment. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8925â8935. PMLR, 18â24 Jul 2021.
- [47] Ravi Francesco Srinivasan, Francesca Mignacco, Martino Sorbaro, Maria Refinetti, Avi Cooper, Gabriel Kreiman, and Giorgia Dellaferrera. Forward learning with top-down feedback: Empirical and analytical characterization. In The Twelfth International Conference on Learning Representations, 2024.
- [48] Nishil Patel, Sebastian Lee, Stefano Sarao Mannelli, Sebastian Goldt, and Andrew Saxe. Rl perceptron: Generalization dynamics of policy learning in high dimensions. Phys. Rev. X, 15:021051, May 2025.
- [49] Tianyi Zhou, Shengjie Wang, and Jeff Bilmes. Curriculum learning by optimizing learning dynamics. In Proceedings of The 24th International Conference on Artificial Intelligence and Statistics, volume 130 of Proceedings of Machine Learning Research, pages 433â441. PMLR, 13â15 Apr 2021.
- [50] LS Pontryagin. Some mathematical problems arising in connection with the theory of optimal automatic control systems. In Proc. Conf. on Basic Problems in Automatic Control and Regulation, 1957.
- [51] Donald E Kirk. Optimal control theory: an introduction. Courier Corporation, 2004.
- [52] John Bechhoefer. Control theory for physicists. Cambridge University Press, 2021.
- [53] John T Betts. Practical methods for optimal control and estimation using nonlinear programming. SIAM, 2010.
- [54] Joel AE Andersson, Joris Gillis, Greg Horn, James B Rawlings, and Moritz Diehl. Casadi: a software framework for nonlinear optimization and optimal control. Mathematical Programming Computation, 11:1â36, 2019.
- [55] Dayal Singh Kalra and Maissam Barkeshli. Why warmup the learning rate? underlying mechanisms and improvements. Advances in Neural Information Processing Systems, 37:111760â111801, 2024.
- [56] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In International Conference on Learning Representations (ICLR), 2017.
- [57] Atilim Gunes Baydin, Robert Cornish, David Martinez Rubio, Mark Schmidt, and Frank Wood. Online learning rate adaptation with hypergradient descent. In International Conference on Learning Representations (ICLR), 2018.
- [58] E Schlösser, D Saad, and M Biehl. Optimization of on-line principal component analysis. Journal of Physics A: Mathematical and General, 32(22):4061, 1999.
- [59] StĂ©phane dâAscoli, Maria Refinetti, and Giulio Biroli. Optimal learning rate schedules in high-dimensional non-convex optimization problems. arXiv preprint arXiv:2202.04509, 2022.
- [60] Lukas Balles, Javier Romero, and Philipp Hennig. Coupling adaptive batch sizes with learning rates. In Proceedings of the Conference on Uncertainty in Artificial Intelligence (UAI), 2017.
- [61] Samuel L Smith, Pieter-Jan Kindermans, Chris Ying, and Quoc V Le. Donât decay the learning rate, increase the batch size. In International Conference on Learning Representations (ICLR), 2018.
- [62] Aditya Devarakonda, Maxim Naumov, and Michael Garland. Adabatch: Adaptive batch sizes for training deep neural networks. In ICLR 2018 Workshop on Optimization for Machine Learning, 2018.
- [63] Huan Wang, Can Qin, Yulun Zhang, and Yun Fu. Neural pruning via growing regularization. In International Conference on Learning Representations (ICLR), 2021.
- [64] David Saad and Magnus Rattray. Learning with regularizers in multilayer neural networks. Physical Review E, 57(2):2170, 1998.
- [65] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation, volume 24, pages 109â165. Elsevier, 1989.
- [66] Ian J. Goodfellow, Mehdi Mirza, Xia Da, Aaron C. Courville, and Yoshua Bengio. An empirical investigation of catastrophic forgeting in gradient-based neural networks. In 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings, 2014.
- [67] Sebastian Lee, Sebastian Goldt, and Andrew Saxe. Continual learning in the teacher-student setup: Impact of task similarity. In International Conference on Machine Learning, pages 6109â6119. PMLR, 2021.
- [68] Sebastian Lee, Stefano Sarao Mannelli, Claudia Clopath, Sebastian Goldt, and Andrew Saxe. Maslowâs hammer in catastrophic forgetting: Node re-use vs. node activation. In International Conference on Machine Learning, pages 12455â12477. PMLR, 2022.
- [69] Itay Evron, Edward Moroshko, Rachel Ward, Nathan Srebro, and Daniel Soudry. How catastrophic can catastrophic forgetting be in linear regression? In Conference on Learning Theory, pages 4028â4079. PMLR, 2022.
- [70] Itay Evron, Edward Moroshko, Gon Buzaglo, Maroun Khriesh, Badea Marjieh, Nathan Srebro, and Daniel Soudry. Continual learning in linear classification on separable data. In International Conference on Machine Learning, pages 9440â9484. PMLR, 2023.
- [71] Haozhe Shan, Qianyi Li, and Haim Sompolinsky. Order parameters and phase transitions of continual learning in deep neural networks. arXiv preprint arXiv:2407.10315, 2024.
- [72] Elisabetta Cornacchia and Elchanan Mossel. A mathematical model for curriculum learning for parities. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 6402â6423. PMLR, 23â29 Jul 2023.
- [73] Emmanuel Abbe, Elisabetta Cornacchia, and Aryo Lotfi. Provable advantage of curriculum learning on parity targets with mixed inputs. In Advances in Neural Information Processing Systems, volume 36, pages 24291â24321. Curran Associates, Inc., 2023.
- [74] Fadi Thabtah, Suhel Hammoud, Firuz Kamalov, and Amanda Gonsalves. Data imbalance in classification: Experimental evaluation. Information Sciences, 513:429â441, 2020.
- [75] Emanuele Loffredo, Mauro Pastore, Simona Cocco, and Remi Monasson. Restoring balance: principled under/oversampling of data for optimal classification. In Forty-first International Conference on Machine Learning, 2024.
- [76] Emanuele Loffredo, Mauro Pastore, Simona Cocco, and RĂ©mi Monasson. Restoring data balance via generative models of t-cell receptors for antigen-binding prediction. bioRxiv, pages 2024â07, 2024.
- [77] Stefano Sarao Mannelli, Federica Gerace, Negar Rostamzadeh, and Luca Saglietti. Bias-inducing geometries: exactly solvable data model with fairness implications. In ICML 2024 Workshop on Geometry-grounded Representation Learning and Generative Modeling, 2024.
- [78] Anchit Jain, Rozhin Nobahari, Aristide Baratin, and Stefano Sarao Mannelli. Bias in motion: Theoretical insights into the dynamics of bias in sgd training. In Advances in Neural Information Processing Systems, volume 37, pages 24435â24471. Curran Associates, Inc., 2024.
- [79] Yizeng Han, Gao Huang, Shiji Song, Le Yang, Honghui Wang, and Yulin Wang. Dynamic neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(11):7436â7456, 2021.
- [80] Geoffrey E Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan R Salakhutdinov. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint arXiv:1207.0580, 2012.
- [81] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929â1958, 2014.
- [82] Pietro Morerio, Jacopo Cavazza, Riccardo Volpi, RenĂ© Vidal, and Vittorio Murino. Curriculum dropout. In Proceedings of the IEEE International Conference on Computer Vision, pages 3544â3552, 2017.
- [83] Zhuang Liu, Zhiqiu Xu, Joseph Jin, Zhiqiang Shen, and Trevor Darrell. Dropout reduces underfitting. In International Conference on Machine Learning, pages 22233â22248. PMLR, 2023.
- [84] Francesco Mori and Francesca Mignacco. Analytic theory of dropout regularization. arXiv preprint arXiv:2505.07792, 2025.
- [85] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In International Conference on Learning Representations, 2017.
- [86] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132â7141, 2018.
- [87] Kyunghyun Cho, Bart Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk, and Y. Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. 06 2014.
- [88] Joel Veness, Tor Lattimore, David Budden, Avishkar Bhoopchand, Christopher Mattern, Agnieszka Grabska-Barwinska, Eren Sezener, Jianan Wang, Peter Toth, Simon Schmitt, et al. Gated linear networks. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 10015â10023, 2021.
- [89] Qianyi Li and Haim Sompolinsky. Globally gated deep linear networks. Advances in Neural Information Processing Systems, 35:34789â34801, 2022.
- [90] Andrew Saxe, Shagun Sodhani, and Sam Jay Lewallen. The neural race reduction: Dynamics of abstraction in gated networks. In International Conference on Machine Learning, pages 19287â19309. PMLR, 2022.
- [91] Samuel Lippl, LF Abbott, and SueYeon Chung. The implicit bias of gradient descent on generalized gated linear networks. arXiv preprint arXiv:2202.02649, 2022.
- [92] Francesca Mignacco, Chi-Ning Chou, and SueYeon Chung. Nonlinear classification of neural manifolds with contextual information. Physical Review E, 111(3):035302, 2025.
- [93] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ćukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.
- [94] Aurko Roy, Mohammad Saffar, Ashish Vaswani, and David Grangier. Efficient content-based sparse attention with routing transformers. Transactions of the Association for Computational Linguistics, 9:53â68, 2021.
- [95] Sainbayar Sukhbaatar, Ădouard Grave, Piotr Bojanowski, and Armand Joulin. Adaptive attention span in transformers. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 331â335, 2019.
- [96] Paul Michel, Omer Levy, and Graham Neubig. Are sixteen heads really better than one? Advances in neural information processing systems, 32, 2019.
- [97] Gonçalo M Correia, Vlad Niculae, and André FT Martins. Adaptively sparse transformers. arXiv preprint arXiv:1909.00015, 2019.
- [98] Hugo Cui, Freya Behrens, Florent Krzakala, and Lenka ZdeborovĂĄ. A phase transition between positional and semantic learning in a solvable model of dot-product attention. Advances in Neural Information Processing Systems, 37:36342â36389, 2024.
- [99] Luca Arnaboldi, Bruno Loureiro, Ludovic Stephan, Florent Krzakala, and Lenka Zdeborova. Asymptotics of sgd in sequence-single index models and single-layer attention networks, 2025.
- [100] Douglas H Lawrence. The transfer of a discrimination along a continuum. Journal of Comparative and Physiological Psychology, 45(6):511, 1952.
- [101] Renee Elio and John R Anderson. The effects of information order and learning mode on schema abstraction. Memory & cognition, 12(1):20â30, 1984.
- [102] Harold Pashler and Michael C Mozer. When does fading enhance perceptual category learning? Journal of Experimental Psychology: Learning, Memory, and Cognition, 39(4):1162, 2013.
- [103] William L Tong, Anisha Iyer, Venkatesh N Murthy, and Gautam Reddy. Adaptive algorithms for shaping behavior. bioRxiv, 2023.
- [104] Yoshua Bengio, JĂ©rĂŽme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41â48, 2009.
- [105] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE transactions on pattern analysis and machine intelligence, 44(9):4555â4576, 2021.
- [106] Anastasia Pentina, Viktoriia Sharmanska, and Christoph H Lampert. Curriculum learning of multiple tasks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5492â5500, 2015.
- [107] Guy Hacohen and Daphna Weinshall. On the power of curriculum learning in training deep networks. In International conference on machine learning, pages 2535â2544. PMLR, 2019.
- [108] Xiaoxia Wu, Ethan Dyer, and Behnam Neyshabur. When do curricula work? In International Conference on Learning Representations (ICLR), 2020.
- [109] Daphna Weinshall and Dan Amir. Theory of curriculum learning, with convex loss functions. Journal of Machine Learning Research, 21(222):1â19, 2020.
- [110] Luca Saglietti, Stefano Mannelli, and Andrew Saxe. An analytical theory of curriculum learning in teacher-student networks. Advances in Neural Information Processing Systems, 35:21113â21127, 2022.
- [111] Stefano Sarao Mannelli, Yaraslau Ivashynka, Andrew Saxe, and Luca Saglietti. Tilting the odds at the lottery: the interplay of overparameterisation and curricula in neural networks. Journal of Statistical Mechanics: Theory and Experiment, 2024(11):114001, 2024.
- [112] Emmanuel Abbe, Elisabetta Cornacchia, and Aryo Lotfi. Provable advantage of curriculum learning on parity targets with mixed inputs. Advances in Neural Information Processing Systems, 36:24291â24321, 2023.
- [113] Elisabetta Cornacchia and Elchanan Mossel. A mathematical model for curriculum learning for parities. In International Conference on Machine Learning, pages 6402â6423. PMLR, 2023.
- [114] Imrus Salehin and Dae-Ki Kang. A review on dropout regularization approaches for deep neural networks within the scholarly domain. Electronics, 12(14):3106, 2023.
- [115] Steven J. Rennie, Vaibhava Goel, and Samuel Thomas. Annealed dropout training of deep networks. In 2014 IEEE Spoken Language Technology Workshop (SLT), pages 159â164, 2014.
- [116] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th International Conference on Machine Learning, ICML â08, page 1096â1103, New York, NY, USA, 2008. Association for Computing Machinery.
- [117] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, and Pierre-Antoine Manzagol. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. J. Mach. Learn. Res., 11:3371â3408, December 2010.
- [118] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning Research, pages 2256â2265, Lille, France, 07â09 Jul 2015. PMLR.
- [119] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, volume 33, pages 6840â6851. Curran Associates, Inc., 2020.
- [120] Arnu Pretorius, Steve Kroon, and Herman Kamper. Learning dynamics of linear denoising autoencoders. In International Conference on Machine Learning, pages 4141â4150. PMLR, 2018.
- [121] Hugo Cui and Lenka ZdeborovĂĄ. High-dimensional asymptotics of denoising autoencoders. Advances in Neural Information Processing Systems, 36:11850â11890, 2023.
- [122] Hugo Cui, Florent Krzakala, Eric Vanden-Eijnden, and Lenka ZdeborovĂĄ. Analysis of learning a flow-based generative model from limited sample complexity. In International Conference on Learning Representations (ICLR), 2024.
- [123] Hugo Cui, Cengiz Pehlevan, and Yue M Lu. A precise asymptotic analysis of learning diffusion models: theory and insights. arXiv preprint arXiv:2501.03937, 2025.
- [124] Maria Refinetti and Sebastian Goldt. The dynamics of representation learning in shallow, non-linear autoencoders. In International Conference on Machine Learning, pages 18499â18519. PMLR, 2022.
- [125] Krzysztof J. Geras and Charles Sutton. Scheduled denoising autoencoders. In International Conference on Learning Representations (ICLR), 2015.
- [126] Tianyi Zheng, Cong Geng, Peng-Tao Jiang, Ben Wan, Hao Zhang, Jinwei Chen, Jia Wang, and Bo Li. Non-uniform timestep sampling: Towards faster diffusion model training. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7036â7045, 2024.
- [127] Minmin Chen, Kilian Weinberger, Fei Sha, and Yoshua Bengio. Marginalized denoising auto-encoders for nonlinear representations. In International conference on machine learning, pages 1476â1484. PMLR, 2014.
- [128] Maria Refinetti, Sebastian Goldt, Florent Krzakala, and Lenka ZdeborovĂĄ. Classifying high-dimensional gaussian mixtures: Where kernel methods fail and neural networks succeed. In International Conference on Machine Learning, pages 8936â8947. PMLR, 2021.
- [129] Sebastian Goldt, Marc Mézard, Florent Krzakala, and Lenka Zdeborovå. Modeling the influence of data structure on learning in neural networks: The hidden manifold model. Physical Review X, 10(4):041044, 2020.
- [130] Sebastian Goldt, Bruno Loureiro, Galen Reeves, Florent Krzakala, Marc MĂ©zard, and Lenka ZdeborovĂĄ. The gaussian equivalence of generative models for learning with shallow neural networks. In Mathematical and Scientific Machine Learning, pages 426â471. PMLR, 2022.
- [131] Yatin Dandi, Emanuele Troiani, Luca Arnaboldi, Luca Pesce, Lenka Zdeborova, and Florent Krzakala. The benefits of reusing batches for gradient descent in two-layer networks: Breaking the curse of information and leap exponents. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 9991â10016. PMLR, 21â27 Jul 2024.
- [132] Andrea Montanari and Pierfrancesco Urbani. Dynamical decoupling of generalization and overfitting in large two-layer networks. arXiv preprint arXiv:2502.21269, 2025.
- [133] Santiago Aranguri, Giulio Biroli, Marc Mezard, and Eric Vanden-Eijnden. Optimizing noise schedules of generative models in high dimensionss. arXiv preprint arXiv:2501.00988, 2025.
- [134] Maria Refinetti, StĂ©phane dâAscoli, Ruben Ohana, and Sebastian Goldt. Align, then memorise: the dynamics of learning with feedback alignment. In International Conference on Machine Learning, pages 8925â8935. PMLR, 2021.
- [135] Blake Bordelon and Cengiz Pehlevan. The influence of learning rule on representation dynamics in wide neural networks. In The Eleventh International Conference on Learning Representations.
- [136] Ravi Francesco Srinivasan, Francesca Mignacco, Martino Sorbaro, Maria Refinetti, Avi Cooper, Gabriel Kreiman, and Giorgia Dellaferrera. Forward learning with top-down feedback: Empirical and analytical characterization. In International Conference on Learning Representations (ICLR), 2024.
## Appendix A Derivation of the learning dynamics
In this section, we derive the set of ordinary differential equations (ODEs) for the order parameters given in Eq. (8) of the main text, that track the dynamics of online stochastic gradient descent (SGD). We consider the cost function
$$
\mathcal{L}({\bm{w}},{\bm{v}}|\bm{x},\bm{c})=\ell\left(\frac{{\bm{x}}^{\top}{
\bm{w}_{*}}}{\sqrt{N}},\frac{{\bm{x}}^{\top}{\bm{w}}}{\sqrt{N}},\frac{\bm{w}^{
\top}\bm{w}}{N},{\bm{v}},{\bm{c}},z\right)+\tilde{g}\left(\frac{\bm{w}^{\top}
\bm{w}}{N},{\bm{v}}\right)\,. \tag{27}
$$
The update rules for the networkâs parameters are
$$
\displaystyle\begin{split}\bm{w}^{\mu+1}=\bm{w}^{\mu}-\eta\nabla_{\bm{w}}
\mathcal{L}(\bm{w}^{\mu},\bm{v}^{\mu}|\bm{x}^{\mu},\bm{c}^{\mu})=\bm{w}^{\mu}-
\eta\left[\frac{{\bm{x}^{\mu}}\nabla_{2}\ell^{\mu}}{\sqrt{N}}+2\frac{\bm{w}^{
\mu}\nabla_{3}\ell^{\mu}}{N}+2\frac{\bm{w}^{\mu}\nabla_{1}\tilde{g}^{\mu}}{N}
\right]\;,\end{split} \displaystyle\begin{split}\bm{v}^{\mu+1}=\bm{v}^{\mu}-\frac{\eta}{N}\nabla_{4}
\ell^{\mu}-\frac{\eta}{N}\nabla_{2}\tilde{g}^{\mu}\;,\end{split} \tag{28}
$$
where we use $\nabla_{k}\ell$ to denote the gradient of the function $\ell$ with respect to its $k^{\rm th}$ argument, with the convention that it is reshaped as a matrix of the same dimensions of that argument, e.g., $\nabla_{2}\ell\in\mathbb{R}^{L\times K}$ . For simplicity, we omit the functionâs arguments, by only keeping the time dependence, i.e., $\ell^{\mu}=\ell\left(\frac{{{\bm{x}}^{\mu}}^{\top}{\bm{w}_{*}}}{\sqrt{N}}, \frac{{{\bm{x}}^{\mu}}^{\top}{\bm{w}}^{\mu}}{\sqrt{N}},\frac{{\bm{w}^{\mu}}^{ \top}\bm{w}^{\mu}}{N},{\bm{v}}^{\mu},{\bm{c}}^{\mu},z^{\mu}\right)$ . For a given realization of the cluster coefficients $\bm{c}$ , we introduce the compact notation $\bm{\mu}_{\bm{c}}\in\mathbb{R}^{N\times L}$ to denote the matrix with columns $\bm{\mu}_{l,c_{l}}$ . It is useful to define the local fields
$$
\displaystyle\bm{\lambda}^{\mu}=\frac{{\bm{x}^{\mu}}^{\top}\bm{w}^{\mu}}{\sqrt
{N}}\in\mathbb{R}^{L\times K}\;, \displaystyle\bm{\lambda}_{*}^{\mu}=\frac{{\bm{x}^{\mu}}^{\top}\bm{w}_{*}}{
\sqrt{N}}\in\mathbb{R}^{L\times M}\;, \displaystyle\bm{\rho}^{\mu}_{\bm{c}}=\frac{{\bm{x}^{\mu}}^{\top}\bm{\mu}_{\bm
{c}}}{\sqrt{N}}\in\mathbb{R}^{L\times L}\;. \tag{30}
$$
Notice that, due to the online-learning setup, at each training step the input $\bm{x}$ is independent of the weights. Therefore, due to the Gaussianity of the inputs, the local fields are also jointly Gaussian with zero mean and second moments given by:
$$
\displaystyle\begin{split}\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{lk}\lambda_
{l^{\prime}k^{\prime}}\right]&=\frac{{\bm{w}_{k}}\cdot\bm{\mu}_{l,c_{l}}}{{N}}
\frac{{\bm{w}_{k^{\prime}}}\cdot\bm{\mu}_{l^{\prime},c_{l^{\prime}}}}{{N}}+
\delta_{l,l^{\prime}}\,\sigma^{2}_{l,c_{l}}\frac{{\bm{w}_{k}}\cdot\bm{w}_{k^{
\prime}}}{N}\\
&=R_{k(l,c_{l})}R_{k^{\prime}(l^{\prime},c_{l^{\prime}})}+\delta_{l,l^{\prime}
}\sigma^{2}_{l,c_{l}}Q_{kk^{\prime}}\;,\end{split} \displaystyle\begin{split}\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{lk}\lambda_
{*,l^{\prime}m}\right]&=\frac{{\bm{w}_{k}}\cdot\bm{\mu}_{l,c_{l}}}{{N}}\frac{{
\bm{w}_{*,m}}\cdot\bm{\mu}_{l^{\prime},c_{l^{\prime}}}}{{N}}+\delta_{l,l^{
\prime}}\,\sigma^{2}_{l,c_{l}}\frac{{\bm{w}_{k}}\cdot\bm{w}_{*,m}}{N}\\
&=R_{k(l,c_{l})}S_{m(l^{\prime},c_{l^{\prime}})}+\delta_{l,l^{\prime}}\sigma^{
2}_{l,c_{l}}M_{km}\;,\end{split} \displaystyle\begin{split}\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{*,lm}
\lambda_{*,l^{\prime}m^{\prime}}\right]&=\frac{{\bm{w}_{*,m}}\cdot\bm{\mu}_{l,
c_{l}}}{{N}}\frac{{\bm{w}_{*,m^{\prime}}}\cdot\bm{\mu}_{l^{\prime},c_{l^{
\prime}}}}{{N}}+\delta_{l,l^{\prime}}\,\sigma^{2}_{l,c_{l}}\frac{{\bm{w}_{*,m}
}\cdot\bm{w}_{*,m^{\prime}}}{N}\\
&=S_{m(l,c_{l})}S_{m^{\prime}(l^{\prime},c_{l^{\prime}})}+\delta_{l,l^{\prime}
}\sigma^{2}_{l,c_{l}}T_{mm^{\prime}}\;,\end{split} \tag{31}
$$
$$
\displaystyle\begin{split}\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{lk}\rho_{{
\bm{c}^{\prime}},l^{\prime}l^{\prime\prime}}\right]&=\frac{{\bm{w}_{k}}\cdot
\bm{\mu}_{l,c_{l}}}{{N}}\frac{\bm{\mu}_{l^{\prime},c_{l^{\prime}}}\cdot\bm{\mu
}_{l^{\prime\prime},c^{\prime}_{l^{\prime\prime}}}}{N}+\delta_{l,l^{\prime}}\,
\sigma^{2}_{l,c_{l}}\frac{{\bm{w}_{k}}\cdot\bm{\mu}_{l^{\prime\prime},c^{
\prime}_{l^{\prime\prime}}}}{N}\\
&=R_{k(l,c_{l})}\Omega_{(l^{\prime},c_{l^{\prime}})(l^{\prime\prime},c^{\prime
}_{l^{\prime\prime}})}+\delta_{l,l^{\prime}}\sigma^{2}_{l,c_{l}}R_{k(l^{\prime
\prime},c^{\prime}_{l^{\prime\prime}})}\;,\end{split} \displaystyle\begin{split}\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{*,lm}\rho_{
{\bm{c}}^{\prime},l^{\prime}l^{\prime\prime}}\right]&=\frac{{\bm{w}_{*,m}}
\cdot\bm{\mu}_{l,c_{l}}}{{N}}\frac{\bm{\mu}_{l^{\prime},c_{l^{\prime}}}\cdot
\bm{\mu}_{l^{\prime\prime},c^{\prime}_{l^{\prime\prime}}}}{N}+\delta_{l,l^{
\prime}}\,\sigma^{2}_{l,c_{l}}\frac{{\bm{w}_{*,m}}\cdot\bm{\mu}_{l^{\prime
\prime},c^{\prime}_{l^{\prime\prime}}}}{N}\\
&=S_{m(l,c_{l})}\Omega_{(l^{\prime},c_{l^{\prime}})(l^{\prime\prime},c^{\prime
}_{l^{\prime\prime}})}+\delta_{l,l^{\prime}}\sigma^{2}_{l,c_{l}}S_{m(l^{\prime
\prime},c^{\prime}_{l^{\prime\prime}})}\;,\end{split} \displaystyle\begin{split}\mathbb{E}_{\bm{x}|\bm{c}}\left[\rho_{{\bm{c}}^{
\prime},ll^{\prime}}\rho_{{\bm{c}}^{\prime\prime},l^{\prime\prime}l^{\prime
\prime\prime}}\right]&=\frac{{\bm{\mu}_{l^{\prime},c^{\prime}_{l^{\prime}}}}
\cdot\bm{\mu}_{l,c_{l}}}{N}\frac{\bm{\mu}_{l^{\prime\prime},c_{l^{\prime\prime
}}}\cdot\bm{\mu}_{l^{\prime\prime\prime},c^{\prime\prime}_{l^{\prime\prime
\prime}}}}{N}+\delta_{l,l^{\prime\prime}}\,\sigma^{2}_{l,c_{l}}\frac{{\bm{\mu}
_{l^{\prime},c^{\prime}_{l^{\prime}}}}\cdot\bm{\mu}_{l^{\prime\prime\prime},c^
{\prime\prime}_{l^{\prime\prime\prime}}}}{N}\\
&=\Omega_{(l,c_{l})(l^{\prime},c^{\prime}_{l^{\prime}})}\Omega_{(l^{\prime
\prime},c_{l^{\prime\prime}})(l^{\prime\prime\prime},c^{\prime\prime}_{l^{
\prime\prime\prime}})}+\delta_{l,l^{\prime\prime}}\sigma^{2}_{l,c_{l}}\Omega_{
(l^{\prime},c^{\prime}_{l^{\prime}})(l^{\prime\prime\prime},c^{\prime\prime}_{
l^{\prime\prime\prime}})}\;,\end{split} \tag{34}
$$
where we have introduced the order parameters
$$
\displaystyle\begin{split}&Q_{kk^{\prime}}\coloneqq\frac{{\bm{w}_{k}}\cdot\bm{
w}_{k^{\prime}}}{N}\;,\quad M_{km}\coloneqq\frac{{\bm{w}^{\mu}_{k}}\cdot\bm{w}
_{*,m}}{N}\;,\quad R_{k(l,c_{l})}\coloneqq\frac{{\bm{w}_{k}}\cdot\bm{\mu}_{l,c
_{l}}}{{N}}\;,\\
&S_{m(l,c_{l})}\coloneqq\frac{{\bm{w}_{*,m}}\cdot\bm{\mu}_{l,c_{l}}}{{N}}\;,
\quad T_{mm^{\prime}}\coloneqq\frac{{\bm{w}_{*,m}}\cdot\bm{w}_{*,m^{\prime}}}{
N}\;,\quad\Omega_{(l,c_{l})(l^{\prime},c^{\prime}_{l^{\prime}})}=\frac{\bm{\mu
}_{l,c_{l}}\cdot\bm{\mu}_{l^{\prime},c^{\prime}_{l^{\prime}}}}{N}\;.\end{split} \tag{37}
$$
Note that in the expressions above the variable $\bm{x}$ is assumed to be drawn from the distribution in Eq. (1) with cluster membership $\bm{c}$ fixed. The additional cluster membership variables, e.g., $\bm{c}^{\prime}$ and $\bm{c}^{\prime\prime}$ are fixed and do not intervene in the generative process of $\bm{x}$ . The cost function defined in Eq. (27) depends on the weights $\bm{w}$ only through the local fields and the order parameters. Similarly, the generalization error (defined in Eq. (7) of the main text) can be computed as an average over the local fields
$$
\displaystyle\varepsilon_{g}(\bm{w},\bm{v})=\mathbb{E}_{\bm{c}}\mathbb{E}_{(
\bm{\lambda},\bm{\lambda}_{*})|\bm{c}}\left[\ell_{g}\left(\bm{\lambda}_{*},\bm
{\lambda},\bm{Q},\bm{v},\bm{c},0\right)\right]\;, \tag{38}
$$
where the function $\ell_{g}$ may coincide with the loss $\ell$ or denote a different metric depending on the context.
Since the local fields are Gaussian, their distribution is completely specified by the first two moments, which are functions of the order parameters. By substituting the update rules of Eq. (28) into the definitions in Eq. (LABEL:eq:orderparams_supmat), we obtain the following evolution equations governing the orderâparameter dynamics
$$
\displaystyle\begin{split}&\bm{Q}^{\mu+1}-\bm{Q}^{\mu}=\frac{{\bm{w}^{\mu+1}}^
{\top}\bm{w}^{\mu+1}}{N}-\frac{{\bm{w}^{\mu}}^{\top}\bm{w}^{\mu}}{N}=\\
&\quad-\frac{\eta}{N}\left[{\bm{\lambda}^{\mu}}^{\top}\nabla_{2}\ell^{\mu}+
\nabla_{2}{\ell^{\mu}}^{\top}{\bm{\lambda}^{\mu}}+2\bm{Q}^{\mu}\left(\nabla_{3
}\ell^{\mu}+\nabla_{1}\tilde{g}^{\mu}\right)+2\left({\nabla_{3}\ell^{\mu}}+
\nabla_{1}\tilde{g}^{\mu}\right)^{\top}\bm{Q}^{\mu}\right]\\
&\quad+\frac{\eta^{2}}{N}\left[{\nabla_{2}\ell^{\mu}}^{\top}\frac{{\bm{x}^{\mu
}}^{\top}{\bm{x}^{\mu}}}{N}\nabla_{2}\ell^{\mu}+\mathcal{O}\left(\frac{1}{N}
\right)\right]\;,\end{split} \displaystyle\begin{split}\bm{M}^{\mu+1}-\bm{M}^{\mu}=\frac{{\bm{w}^{\mu+1}}^{
\top}\bm{w}_{*}}{N}-\frac{{\bm{w}^{\mu}}^{\top}\bm{w}_{*}}{N}=-\frac{\eta}{N}
\left[{\nabla_{2}\ell^{\mu}}^{\top}\bm{\lambda}_{*}^{\mu}+2\left(\nabla_{3}
\ell^{\mu}+\nabla_{1}\tilde{g}^{\mu}\right)^{\top}\bm{M}^{\mu}\right]\;,\end{split} \displaystyle\begin{split}\bm{R}_{\bm{c}^{\prime}}^{\mu+1}-\bm{R}_{\bm{c}^{
\prime}}^{\mu}=\frac{{\bm{w}^{\mu+1}}^{\top}\bm{\mu}_{\bm{c}^{\prime}}}{{N}}-
\frac{{\bm{w}^{\mu}}^{\top}\bm{\mu}_{\bm{c}^{\prime}}}{{N}}=-\frac{\eta}{N}
\left[{\nabla_{2}\ell^{\mu}}^{\top}{\bm{\rho}}_{\bm{c}^{\prime}}+2\left(\nabla
_{3}\ell^{\mu}+\nabla_{1}\tilde{g}\right)^{\top}\bm{R}_{\bm{c}^{\prime}}^{\mu}
\right]\;,\end{split} \tag{39}
$$
where we have omitted subleading terms in $N$ . Note that, while for convenience we write $\bm{R}_{\bm{c}^{\prime}}$ for an arbitrary cluster membership variable ${\bm{c}}^{\prime}=(c^{\prime}_{1}\,,\ldots\,,c^{\prime}_{L})$ , it is sufficient to keep track of the scalar variables $R_{k},(l,c^{\prime\prime}_{l})$ for $k=1\,,\ldots K$ , $l=1\,,\ldots\,,L$ , $c^{\prime\prime}_{l}=1\,,\ldots\,,C_{l}$ , resulting in $K(C_{1}+C_{2}+\ldots+C_{L})$ variables. We define a âtraining timeâ $\alpha=\mu/N$ and take the infinite-dimensional limit $N\rightarrow\infty$ while keeping $\alpha$ of order one. We obtain the following ODEs
$$
\displaystyle\begin{split}\frac{{\rm d}\bm{Q}}{{\rm d}\alpha}&=\mathbb{E}_{\bm
{c}}\Big{[}-\eta\left\{\mathbb{E}_{\bm{\lambda},\bm{\lambda}_{*}|\bm{c}}\left[
\bm{\lambda}^{\top}\nabla_{2}\ell\right]+2\,\bm{Q}\left(\mathbb{E}_{\bm{
\lambda},\bm{\lambda}_{*}|\bm{c}}\left[\nabla_{3}\ell\right]+\nabla_{1}\tilde{
g}\right)+{\rm(transpose)}\right\}\\
&\qquad\qquad+\eta^{2}\,\mathbb{E}_{\bm{\lambda},\bm{\lambda}_{*}|\bm{c}}\left
[\nabla_{2}\ell^{\top}{\rm diag}(\bm{\sigma^{2}}_{\bm{c}})\nabla_{2}\ell\right
]\Big{]}\coloneqq f_{\bm{Q}}\;,\end{split} \displaystyle\begin{split}\frac{{\rm d}\bm{M}}{{\rm d}\alpha}=\mathbb{E}_{\bm{
c}}\Big{[}-\eta\,\mathbb{E}_{\bm{\lambda},\bm{\lambda}_{*}|\bm{c}}\left[{
\nabla_{2}\ell}^{\top}\bm{\lambda}_{*}\right]-2\eta\left(\mathbb{E}_{\bm{
\lambda},\bm{\lambda}_{*}|\bm{c}}\left[{\nabla_{3}\ell}\right]+\nabla_{1}
\tilde{g}\right)^{\top}\bm{M}\Big{]}\coloneqq f_{\bm{M}}\;,\end{split} \displaystyle\begin{split}\frac{{\rm d}\bm{R}_{\bm{c}^{\prime}}}{{\rm d}\alpha
}=\mathbb{E}_{\bm{c}}\Big{[}-\eta\,\mathbb{E}_{\bm{\lambda},\bm{\lambda}_{*}|
\bm{c}}\left[\nabla_{2}\ell^{\top}\bm{\rho}_{\bm{c}^{\prime}}\right]-2\eta
\left(\mathbb{E}_{\bm{\lambda},\bm{\lambda}_{*}|\bm{c}}\left[\nabla_{3}\ell
\right]+\nabla_{1}\tilde{g}\right)^{\top}\bm{R}_{\bm{c}^{\prime}}\Big{]}
\coloneqq f_{\bm{R}_{\bm{c}^{\prime}}}\;,\end{split} \tag{42}
$$
where we remind that $\ell=\ell\left(\bm{\lambda}_{*},\bm{\lambda},\bm{Q},\bm{v},\bm{c},z\right)$ and $\tilde{g}=\tilde{g}(\bm{Q},\bm{v})$ , and we have defined the vector of variances $\bm{\sigma^{2}}_{\bm{c}}=(\sigma^{2}_{1,c_{1}},\ldots,\sigma^{2}_{L,c_{L}})$ . In going from Eq. (LABEL:eq:supmat_evol_Q) to Eq. (42), we have used
$$
\lim_{N\to\infty}\frac{\bm{x}_{l}\cdot\bm{x}_{l^{\prime}}}{N}=\sigma_{l,c_{l}}
^{2}\delta_{ll^{\prime}}\,. \tag{45}
$$
Crucially, when taking the thermodynamic limit $N\to\infty$ , we have replaced the right-hand sides in Eqs. (42)-(44) with their expected value over the data distribution. Indeed, it can be shown rigourously that, under additional assumptions, the fluctuations of the order parameters can be neglected [24]. Although we do not provide a rigorous proof of this result here, we verify this concentration property with numerical simulations, see Appendix C. Finally, the additional parameters $\bm{v}$ evolve according to the low-dimensional equations
$$
\displaystyle\frac{{\rm d}\bm{v}}{{\rm d}\alpha}=\mathbb{E}_{\bm{c}}\Big{[}-
\eta\,\mathbb{E}_{\bm{\lambda},\bm{\lambda}_{*}|\bm{c}}\left[\nabla_{4}\ell+
\nabla_{2}\tilde{g}\right]\Big{]}\coloneqq f_{\bm{v}}\;. \tag{46}
$$
To conclude, note that the expectations in Eqs. (42)â(44) and (46) decompose into an average over the lowâdimensional cluster vector $\mathbf{c}$ , whose distribution is given by the model, and an average over the Gaussian fields $\bm{\lambda}$ and $\bm{\lambda}_{*}$ , whose moments are fully specified by the order parameters, resulting in a closed-form system of equations. The expectations can be evaluated either analytically or via Monte Carlo sampling.
### A.1 Curriculum learning
The equations for the curriculum learning problem can be derived as a special case of those of [110]. The misclassification error can be expressed in terms of the order parameters as
$$
\displaystyle\epsilon_{g}(\mathbb{Q})=\frac{1}{2}-\frac{1}{\pi}\sin^{-1}\left(
\frac{M_{11}}{\sqrt{T(Q_{11}+\Delta Q_{22})}}\right)\;. \tag{47}
$$
The evolution equations for the order parameters can be obtained from Eq. (42) and (44), yielding
$$
\displaystyle\begin{split}\frac{{\rm d}Q_{11}}{{\rm d}\alpha}&=-\bar{\lambda}Q
_{11}+\frac{4\eta}{\pi(Q_{11}+\Delta Q_{22}+2)}\left[\frac{M_{11}(\Delta Q_{22
}+2)}{\sqrt{T(Q_{11}+\Delta Q_{22}+2)-M_{11}^{2}}}-\frac{Q_{11}}{\sqrt{Q_{11}+
\Delta Q_{22}+1}}\right]\\
&\qquad+\frac{2}{\pi^{2}}\frac{\eta^{2}}{\sqrt{Q_{11}+\Delta Q_{22}+1}}\left[
\frac{\pi}{2}+\sin^{-1}\left(\frac{Q_{11}+\Delta Q_{22}}{2+3(Q_{11}+\Delta Q_{
22})}\right)\right.\\
&\qquad\left.-2\sin^{-1}\left(\frac{M_{11}}{\sqrt{\left(3(Q_{11}+\Delta Q_{22}
)+2\right)}\sqrt{T(Q_{11}+\Delta Q_{22}+1)-M_{11}^{2}}}\right)\right]\,,\\
\frac{{\rm d}Q_{22}}{{\rm d}\alpha}&=-\bar{\lambda}Q_{22}-\frac{4\eta\Delta Q_
{22}}{\pi(Q_{11}+\Delta Q_{22}+2)}\left[\frac{M_{11}}{\sqrt{T(Q_{11}+\Delta Q_
{22}+2)-M_{11}^{2}}}+\frac{1}{\sqrt{Q_{11}+\Delta Q_{22}+1}}\right]\\
&\qquad+\frac{2}{\pi^{2}}\frac{\Delta\eta^{2}}{\sqrt{Q_{11}+\Delta Q_{22}+1}}
\left[\frac{\pi}{2}+\sin^{-1}\left(\frac{Q_{11}+\Delta Q_{22}}{2+3(Q_{11}+
\Delta Q_{22})}\right)\right.\\
&\qquad\left.-2\sin^{-1}\left(\frac{M_{11}}{\sqrt{\left(3(Q_{11}+\Delta Q_{22}
)+2\right)}\sqrt{T(Q_{11}+\Delta Q_{22}+1)-M_{11}^{2}}}\right)\right]\,,\\
\frac{{\rm d}M_{11}}{{\rm d}\alpha}&=-\frac{\bar{\lambda}}{2}M_{11}+\frac{2
\eta}{\pi(Q_{11}+\Delta Q_{22}+2)}\left[\sqrt{T(Q_{11}+\Delta Q_{22}+2)-M_{11}
^{2}}-\frac{M_{11}}{\sqrt{Q_{11}+\Delta Q_{22}+1}}\right]\,,\end{split} \tag{48}
$$
where $\bar{\lambda}=\lambda\eta$ .
### A.2 Dropout regularization
In this section, we provide the expressions of the ODEs and the generalization error for the model of dropout regularization presented in Sec. 3.2. This model corresponds to $L=C_{1}=1$ , $\bm{\mu}_{1,1}=\bm{0}$ , and $\sigma_{1,1}=1$ . The derivation of these results can be found in [38]. The generalization error reads
$$
\displaystyle\begin{split}\epsilon_{g}&=\mathbb{E}_{\bm{x}}\left[\frac{1}{2}
\left(f^{*}_{\bm{w}_{*}}(\bm{x})-f^{\rm test}_{\bm{w}}(\bm{x})\right)^{2}
\right]=\frac{p_{f}^{2}}{\pi}\sum_{i,k=1}^{K}\arcsin\left(\frac{Q_{ik}}{\sqrt{
1+Q_{ii}}\sqrt{1+Q_{kk}}}\right)\\
&\quad+\frac{1}{\pi}\sum_{n,m=1}^{K}\arcsin\left(\frac{T_{nm}}{\sqrt{1+T_{nn}}
\sqrt{1+T_{mm}}}\right)-\frac{2p_{f}}{\pi}\sum_{i=1}^{K}\sum_{n=1}^{M}\arcsin
\left(\frac{M_{in}}{\sqrt{1+Q_{ii}}\sqrt{1+T_{nn}}}\right).\end{split} \tag{49}
$$
The ODEs read
$$
\displaystyle\frac{\mathrm{d}M_{in}}{\mathrm{d}\alpha}=f_{M_{in}}(Q,M), \displaystyle\frac{\mathrm{d}Q_{ik}}{\mathrm{d}\alpha}=f_{Q_{ik}}(Q,M), \tag{50}
$$
Introducing the notation
$$
\mathcal{N}\left[r,\{i,j,k,\ldots,l\}\right]=r^{n}\,, \tag{51}
$$
where $n=|\{i,j,k,\ldots,l\}|$ is the cardinality of the set $\{i,j,k,\ldots,l\}$ , we find [84]
$$
\displaystyle f_{M_{in}} \displaystyle\equiv\eta\left[\sum_{m=1}^{M}\mathcal{N}\left[r,\{i\}\right]I_{3
}(i,n,m)-\sum_{j=1}^{K}\mathcal{N}\left[r,\{i,j\}\right]I_{3}(i,n,j)\right], \displaystyle f_{Q_{ik}} \displaystyle\equiv\eta\left[\sum_{m=1}^{M}\mathcal{N}\left[r,\{i\}\right]I_{3
}(i,k,m)-\sum_{j=1}^{K}\mathcal{N}\left[r,\{i,j\}\right]I_{3}(i,k,j)\right] \displaystyle\quad+\eta\left[\sum_{m=1}^{M}\mathcal{N}\left[r,\{k\}\right]I_{3
}(k,i,m)-\sum_{j=1}^{K}\mathcal{N}\left[r,\{k,j\}\right]I_{3}(k,i,j)\right] \displaystyle\quad+\eta^{2}\Bigg{[}\sum_{n=1}^{M}\sum_{m=1}^{M}\mathcal{N}
\left[r,\{i,k\}\right]I_{4}(i,k,n,m)-2\sum_{j=1}^{K}\sum_{n=1}^{M}\mathcal{N}
\left[r,\{i,k,j\}\right]I_{4}(i,k,j,n) \displaystyle\quad\quad+\sum_{j=1}^{K}\sum_{l=1}^{K}\mathcal{N}\left[r,\{i,j,k
,l\}\right]I_{4}(i,k,j,l)+\mathcal{N}\left[r,\{i,k\}\right]\sigma^{2}J_{2}(i,k
)\Bigg{]}, \tag{52}
$$
where
$$
\displaystyle J_{2} \displaystyle\equiv\frac{2}{\pi}\left(1+c_{11}+c_{22}+c_{11}c_{22}-c_{12}^{2}
\right)^{-1/2}, \displaystyle I_{2} \displaystyle\equiv\frac{1}{\pi}\arcsin\left(\frac{c_{12}}{\sqrt{1+c_{11}}
\sqrt{1+c_{12}}}\right), \displaystyle I_{3} \displaystyle\equiv\frac{2}{\pi}\frac{1}{\sqrt{\Lambda_{3}}}\frac{c_{23}(1+c_{
11})-c_{12}c_{13}}{1+c_{11}}, \displaystyle I_{4} \displaystyle\equiv\frac{4}{\pi^{2}}\frac{1}{\sqrt{\Lambda_{4}}}\arcsin\left(
\frac{\Lambda_{0}}{\sqrt{\Lambda_{1}\Lambda_{2}}}\right), \tag{54}
$$
and
$$
\displaystyle\Lambda_{4} \displaystyle=(1+c_{11})(1+c_{22})-c_{12}^{2}, \displaystyle\Lambda_{3} \displaystyle=(1+c_{11})*(1+c_{33})-c_{13}^{2}\,, \displaystyle\Lambda_{0} \displaystyle=\Lambda_{4}c_{34}-c_{23}c_{24}(1+c_{11})-c_{13}c_{14}(1+c_{22})+
c_{12}c_{13}c_{24}+c_{12}c_{14}c_{23}, \displaystyle\Lambda_{1} \displaystyle=\Lambda_{4}(1+c_{33})-c_{23}^{2}(1+c_{11})-c_{13}^{2}(1+c_{22})+
2c_{12}c_{13}c_{23}, \displaystyle\Lambda_{2} \displaystyle=\Lambda_{4}(1+c_{44})-c_{24}^{2}(1+c_{11})-c_{14}^{2}(1+c_{22})+
2c_{12}c_{14}c_{24}. \tag{58}
$$
The indices $i,j,k,l$ and $n,m$ indicate the studentâs and the teacherâs nodes, respectively. For compactness, we adopt the notation for $I_{2}$ , $I_{3}$ , and $I_{4}$ of Ref. [24]. As an example, $I(i,n)$ takes as input the correlation matrix of the preactivations corresponding to the indices $i$ and $n$ , i.e., $\lambda_{i}={\bm{w}}_{i}\cdot{\bm{x}}/\sqrt{N}$ and $\lambda_{*,n}={\bm{w}}^{*}_{n}\cdot{\bm{x}}/\sqrt{N}$ . For this example, the correlation matrix would be
$$
C=\begin{pmatrix}c_{11}&c_{12}\\
c_{21}&c_{22}\end{pmatrix}=\begin{pmatrix}\langle\lambda_{i}\lambda_{i}\rangle
&\langle\lambda_{i}\lambda_{*,n}\rangle\\
\langle\lambda_{*,n}\lambda_{i}\rangle&\langle\lambda_{*,n}\lambda_{*,n}
\rangle\end{pmatrix}=\begin{pmatrix}Q_{ii}&M_{in}\\
M_{in}&T_{nn}\end{pmatrix}\,. \tag{63}
$$
### A.3 Denoising autoencoder
We define the additional local fields
$$
\displaystyle\tilde{\lambda}_{k}\equiv\frac{{\tilde{\bm{x}}}\cdot{\bm{w}}_{k}}
{\sqrt{N}}=\sqrt{1-\Delta}\lambda_{1,k}+\sqrt{\Delta}\lambda_{2,k}\,,\quad
\tilde{\rho}_{{\bm{c}},l}\equiv\frac{{\tilde{\bm{x}}}\cdot{\bm{\mu}}_{l,c_{l}}
}{\sqrt{N}}=\sqrt{1-\Delta}\rho_{{\bm{c}},1l}+\sqrt{\Delta}\rho_{{\bm{c}},2l}\,, \tag{64}
$$
where we recall $\lambda_{1,k}={\bm{w}}_{k}\cdot{\bm{x}}_{1}/\sqrt{N}$ , $\lambda_{2,k}={\bm{w}}_{k}\cdot{\bm{x}}_{2}/\sqrt{N}$ , $\rho_{{\bm{c}},1l}={\bm{\mu}}_{l,c_{l}}\cdot{\bm{x}}_{1}/\sqrt{N}$ , $\rho_{{\bm{c}},2l}={\bm{\mu}}_{l,c_{l}}\cdot{\bm{x}}_{2}/\sqrt{N}$ . Here, we take $C_{2}=1$ and $\bm{\mu}_{2,c_{2}}={\bm{0}}$ , so that $\rho_{{\bm{c}},12}=\rho_{{\bm{c}},22}=\tilde{\rho}_{{\bm{c}},2}=0$ . The local fields are Gaussian variables with moments given by
$$
\displaystyle\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{1,k}\right]=\frac{{\bm{w
}}_{k}\cdot{\bm{\mu}}_{1,c_{1}}}{N}=R_{k(1,c_{1})}\,, \displaystyle\mathbb{E}_{\bm{x}|\bm{c}}\left[\rho_{{\bm{c}^{\prime}},11}\right
]=\frac{\bm{\mu}_{1,c_{1}}\cdot\bm{\mu}_{1,c^{\prime}_{1}}}{N}=\Omega_{(1,c_{1
})(1,c^{\prime}_{1})}\,, \tag{65}
$$
$$
\displaystyle\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{2,k}\right] \displaystyle=\mathbb{E}_{\bm{x}|\bm{c}}\left[\rho_{{\bm{c}^{\prime}},2l}
\right]=0\;, \tag{66}
$$
$$
\displaystyle\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{1,k}\lambda_{2,h}\right]
=\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{1,k}\rho_{{\bm{c}^{\prime}},2l}
\right]=\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{2,k}\rho_{{\bm{c}^{\prime}},1
l}\right]=\mathbb{E}_{\bm{x}|\bm{c}}\left[\rho_{{\bm{c}^{\prime}},1l}\rho_{{
\bm{c}^{\prime}},2l^{\prime}}\right]=0\,, \tag{67}
$$
$$
\displaystyle\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{1,k}\lambda_{1,h}\right]
=R_{k(1,c_{1})}R_{h(1,c1)}+\sigma^{2}_{1,c_{1}}Q_{kh}\,, \displaystyle\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{2,k}\lambda_{2,h}\right]
=Q_{kh}\,, \tag{68}
$$
$$
\displaystyle\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}_{j}\lambda_{1,k}
\right]=\sqrt{1-\Delta}\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{1,k}\lambda_{1
,j}\right]\,, \displaystyle\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}_{j}\lambda_{2,k}
\right]=\sqrt{\Delta}\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{2,k}\lambda_{2,j
}\right]\,, \tag{69}
$$
$$
\displaystyle\mathbb{E}_{\bm{x}|\bm{c}}\left[\rho_{{\bm{c}}^{\prime},11}^{2}
\right]=\Omega_{(1,c_{1})(1,c^{\prime}_{1})}^{2}+\sigma^{2}_{1,c_{1}}\Omega_{(
1,c^{\prime}_{1})(1,c_{1})}\,, \displaystyle\mathbb{E}_{\bm{x}|\bm{c}}\left[\rho_{{\bm{c}^{\prime}},21}^{2}
\right]=\Omega_{(1,c^{\prime}_{1})(1,c^{\prime}_{1})}\,. \tag{70}
$$
$$
\displaystyle\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{1,k}\rho_{{\bm{c}^{
\prime}},11}\right]=\sigma_{1,c_{1}}^{2}R_{k(1,c^{\prime}_{1})}+\Omega_{(1,c^{
\prime}_{1})(1,c_{1})}R_{k(1,c_{1})}\,, \displaystyle\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{2,k}\rho_{{\bm{c}^{
\prime}},21}\right]=R_{k(1,c^{\prime}_{1})}\,. \tag{71}
$$
It is also useful to compute the first moments of the combined variables
$$
\displaystyle\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}_{k}\right]=\sqrt{
1-\Delta}\,R_{k(1,c_{1})}\;, \displaystyle\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\rho}_{{\bm{c}^{\prime}},1
}\right]=\sqrt{1-\Delta}\,\Omega_{(1,c_{1})(1,c^{\prime}_{1})}\,, \tag{72}
$$
and the second moments
$$
\displaystyle\begin{split}\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}_{k}
\tilde{\lambda}_{h}\right]-\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}_{k}
\right]\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}_{h}\right]&=\left[(1-
\Delta)\sigma_{1,c_{1}}^{2}+\Delta\right]Q_{kh}\,,\\
\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\rho}_{{\bm{c}^{\prime}},1}^{2}\right]-
\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\rho}_{{\bm{c}^{\prime}},1}\right]^{2}&
=\left[(1-\Delta)\sigma_{1,c_{1}}^{2}+\Delta\right]\Omega_{(1,c^{\prime}_{1})(
1,c^{\prime}_{1})}\,.\end{split} \tag{73}
$$
Finally, we have
$$
\displaystyle\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}_{k}\rho_{{\bm{c}^
{\prime}},11}\right] \displaystyle=\sqrt{1-\Delta}\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{1,k}\rho
_{{\bm{c}^{\prime}},11}\right]\,, \displaystyle\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}_{k}\tilde{\rho}_{
{\bm{c}^{\prime}},1}\right] \displaystyle=(1-\Delta)\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{1,k}\rho_{{
\bm{c}^{\prime}},11}\right]+\Delta\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{2,k
}\rho_{{\bm{c}^{\prime}},21}\right]\,. \tag{74}
$$
The mean squared error (MSE) can be expressed in terms of the order parameter as follows
$$
\displaystyle\begin{split}\text{MSE}(\bm{w},b)&=\mathbb{E}_{\bm{x},\bm{c}}
\left[\|\bm{x}-f_{\bm{w},b}(\tilde{\bm{x}})\|_{2}^{2}\right]=\mathbb{E}_{\bm{c
}}\left\{N\left[\sigma_{k}^{2}\left(1-b\sqrt{1-\Delta}\right)^{2}+b^{2}\Delta
\right]\right.\\
&\quad+\left.\sum_{j,k=1}^{K}Q_{jk}\mathbb{E}_{\bm{x}|\bm{c}}\left[g(\tilde{
\lambda}_{j})g(\tilde{\lambda}_{k})\right]-2\sum_{k=1}^{K}\mathbb{E}_{\bm{x}|
\bm{c}}\left[(\lambda_{1k}-b\tilde{\lambda}_{k})g(\tilde{\lambda}_{k})\right]
\right\}\,,\end{split} \tag{76}
$$
where we have neglected constant terms. The weights are updated according to
$$
\displaystyle\begin{split}\bm{w}^{\mu+1}_{k}&=\bm{w}^{\mu}_{k}+\frac{\eta}{
\sqrt{N}}g\left(\tilde{\lambda}^{\mu}_{k}\right)\left(\bm{x}_{1}^{\mu}-b\,
\tilde{\bm{x}}^{\mu}-\sum_{h=1}^{K}\frac{{\bm{w}_{h}^{\mu}}}{\sqrt{N}}g\left(
\tilde{\lambda}^{\mu}_{h}\right)\right)\\
&\quad+\frac{\eta}{\sqrt{N}}g^{\prime}(\tilde{\lambda}^{\mu}_{k})\,\left(
\lambda^{\mu}_{1,k}-b\,\tilde{\lambda}^{\mu}_{k}-\sum_{h=1}^{K}\frac{\bm{w}^{
\mu}_{k}\cdot{\bm{w}^{\mu}_{h}}}{{N}}g\left(\tilde{\lambda}^{\mu}_{h}\right)
\right)\,{\tilde{\bm{x}}^{\mu}}\;,\end{split} \tag{77}
$$
The skip connection is also trained with SGD. To leading order, we find
$$
\displaystyle b^{\mu+1}=b^{\mu}+\frac{\eta_{b}}{N}\left(\sqrt{1-\Delta}\sigma_
{1,c_{1}}^{2}-b^{\mu}(1-\Delta)\sigma_{1,c_{1}}^{2}-b^{\mu}\Delta\right)\;. \tag{78}
$$
Note that, conditioning on a given cluster $c_{1}$ , for large $N$ , we have
$$
\displaystyle\frac{1}{N}{\bm{x}_{1}\cdot\bm{x}_{1}}\underset{N\gg 1}{\approx}
\sigma_{1,c_{1}}^{2}\,,\quad\frac{1}{N}{\tilde{\bm{x}}\cdot\tilde{\bm{x}}}
\underset{N\gg 1}{\approx}(1-\Delta)\sigma_{1,c_{1}}^{2}+\Delta\,,\quad\frac{1
}{N}{\bm{x}_{1}\cdot\tilde{\bm{x}}}\underset{N\gg 1}{\approx}\sqrt{1-\Delta}\,
\sigma_{1,c_{1}}^{2}\,. \tag{79}
$$
For simplicity, we will consider the linear activation $g(z)=z$ . In this case, it is possible to derive explicit equations for the evolution of the order parameters as follows:
$$
\displaystyle\begin{split}R^{\mu+1}_{k(1,c^{\prime}_{1})}&=R^{\mu}_{k(1,c^{
\prime}_{1})}+\frac{\eta}{{N}}\mathbb{E}_{\bm{c}}\left[\mathbb{E}_{\bm{x}|\bm{
c}}\left[\tilde{\lambda}^{\mu}_{k}\rho^{\mu}_{\bm{c^{\prime}},11}\right]-2b
\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}^{\mu}_{k}\tilde{\rho}^{\mu}_{
\bm{c^{\prime}},1}\right]-\sum_{j=1}^{K}R^{\mu}_{j(1,c^{\prime}_{1})}\mathbb{E
}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}^{\mu}_{k}\tilde{\lambda}^{\mu}_{j}
\right]\right.\\
&\quad+\left.\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda^{\mu}_{1,k}\tilde{\rho}^{
\mu}_{\bm{c^{\prime}},1}\right]-\sum_{j=1}^{K}Q_{jk}\mathbb{E}_{\bm{x}|\bm{c}}
\left[\tilde{\lambda}^{\mu}_{j}\tilde{\rho}_{\bm{c^{\prime}},1}\right]\right]
\;,\end{split} \tag{80}
$$
$$
\displaystyle\begin{split}Q^{\mu+1}_{jk}&=Q^{\mu}_{jk}+\frac{\eta}{N}\mathbb{E
}_{\bm{c}}\left\{\left(\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}_{j}
\Lambda_{k}\right]+\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}_{k}\Lambda_
{j}\right]\right)\left[2+\eta\left(\frac{\bm{x}_{1}\cdot\tilde{\bm{x}}}{N}-b
\frac{\tilde{\bm{x}}\cdot\tilde{\bm{x}}}{N}\right)\right]+\eta\mathbb{E}_{\bm{
x}|\bm{c}}\left[\Lambda_{j}\Lambda_{k}\right]\frac{\tilde{\bm{x}}\cdot\tilde{
\bm{x}}}{N}\right.\\
&\quad+\left.\eta\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}_{j}\tilde{
\lambda}_{k}\right]\left(\frac{\bm{x}_{1}\cdot\bm{x}_{1}}{N}-2b\frac{\bm{x}_{1
}\cdot\tilde{\bm{x}}}{N}+b^{2}\frac{\tilde{\bm{x}}\cdot\tilde{\bm{x}}}{N}
\right)\right\}\\
&=Q^{\mu}_{jk}+\frac{\eta}{N}\mathbb{E}_{\bm{c}}\left\{\left(\mathbb{E}_{\bm{x
}|\bm{c}}\left[\tilde{\lambda}_{j}\Lambda_{k}\right]+\mathbb{E}_{\bm{x}|\bm{c}
}\left[\tilde{\lambda}_{k}\Lambda_{j}\right]\right)\left[2+\eta\left(\sqrt{1-
\Delta}\sigma_{1,c_{1}}^{2}-b((1-\Delta)\sigma_{1,c_{1}}^{2}+\Delta))\right)
\right]\right.\\
&+\eta\mathbb{E}_{\bm{x}|\bm{c}}\left[\Lambda_{j}\Lambda_{k}\right]((1-\Delta)
\sigma_{1,c_{1}}^{2}+\Delta)+\left.\eta\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{
\lambda}_{j}\tilde{\lambda}_{k}\right]\left(\sigma_{1,c_{1}}^{2}-2b\sqrt{1-
\Delta}\,\sigma_{1,c_{1}}^{2}+b^{2}((1-\Delta)\sigma_{1,c_{1}}^{2}+\Delta))
\right)\right\}\end{split} \tag{81}
$$
where we have introduced the definition
$$
\displaystyle\Lambda_{k}\equiv\lambda_{1,k}-b\tilde{\lambda}_{k}-\sum_{j=1}^{K
}Q_{jk}\tilde{\lambda}_{j}\;. \tag{82}
$$
We can compute the averages
$$
\displaystyle\begin{split}\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}_{j}
\Lambda_{k}\right]&=\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}_{j}\lambda
_{1,k}\right]-\sum_{i=1}^{K}\left(b\delta_{ik}+Q_{ki}\right)\mathbb{E}_{\bm{x}
|\bm{c}}\left[\tilde{\lambda}_{j}\tilde{\lambda}_{i}\right]\;,\\
\mathbb{E}_{\bm{x}|\bm{c}}\left[\Lambda_{j}\Lambda_{k}\right]&=\mathbb{E}_{\bm
{x}|\bm{c}}\left[\lambda_{1,j}\lambda_{1,k}\right]-\sum_{i=1}^{K}\left(b\delta
_{ij}+Q_{ji}\right)\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}_{i}\lambda_
{1,k}\right]-\sum_{i=1}^{K}\left(b\delta_{ik}+Q_{ki}\right)\mathbb{E}_{\bm{x}|
\bm{c}}\left[\tilde{\lambda}_{i}\lambda_{1,j}\right]\\
&\quad+\sum_{i,\ell=1}^{K}\left(b\delta_{ik}+Q_{ki}\right)\left(b\delta_{\ell j
}+Q_{j\ell}\right)\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}_{i}\tilde{
\lambda}_{\ell}\right]\;.\end{split} \tag{83}
$$
Finally, it is useful to evaluate the MSE in the special case of linear activation:
$$
\displaystyle\begin{split}\text{MSE}&=\mathbb{E}_{\bm{c}}\left\{N\left[\sigma_
{1,c_{1}}^{2}\left(1-b\sqrt{1-\Delta}\right)^{2}+b^{2}\Delta\right]\right.\\
&+\sum_{j,k=1}^{K}Q_{jk}\left[\left((1-\Delta)\sigma_{1,c_{1}}^{2}+\Delta
\right)Q_{jk}+(1-\Delta)R_{j,(1,c_{1})}R_{k,(1,c_{1})}\right]\\
&-2\left.\sum_{k=1}^{K}\left[\sqrt{1-\Delta}\sigma_{1,c_{1}}^{2}Q_{kk}-b\left[
\left((1-\Delta)\sigma_{1,c_{1}}^{2}+\Delta\right)Q_{kk}+(1-\Delta)R_{k,(1,c_{
1})}^{2}\right]\right]\right\}\;.\end{split} \tag{84}
$$
#### A.3.1 Data augmentation
We consider inputs $\bm{x}=(\bm{x}_{1},\bm{x}_{2},\ldots,\bm{x}_{B+1})\in\mathbb{R}^{N\times B+1}$ , where $\bm{x}_{1}\sim\mathcal{N}\left(\frac{\bm{\mu}_{1,c_{1}}}{\sqrt{N}},\sigma^{2} \bm{I}_{N}\right)$ denotes the clean input and $\bm{x}_{2},\ldots,\bm{x}_{B+1}\overset{\rm i.i.d.}{\sim}\mathcal{N}(\bm{0},\bm {I}_{N})$ . Each clean input $\bm{x}_{1}$ is used to create multiple corrupted samples: $\tilde{\bm{x}}_{a}=\sqrt{1-\Delta}\,\bm{x}_{1}+\sqrt{\Delta}\,\bm{x}_{a+1}$ , $a=1,\ldots,B$ , that are used as a mini-batch for training. The SGD dynamics of the tied weights modifies as follows:
$$
\displaystyle\begin{split}\bm{w}^{\mu+1}_{k}=\\
\bm{w}^{\mu}_{k}+\frac{\eta}{B^{\mu}\sqrt{N}}\sum_{a=1}^{B^{\mu}}\left\{\tilde
{\lambda}^{\mu}_{a,k}\left(\bm{x}_{1}^{\mu}-b\,\tilde{\bm{x}}^{\mu}_{a}-\sum_{
j=1}^{K}\frac{{\bm{w}^{\mu}_{j}}}{\sqrt{N}}\tilde{\lambda}^{\mu}_{a,j}\right)+
\left(\lambda^{\mu}_{1,k}-b\,\tilde{\lambda}^{\mu}_{a,k}-\sum_{j=1}^{K}\frac{
\bm{w}^{\mu}_{k}\cdot{\bm{w}^{\mu}_{j}}}{{N}}\tilde{\lambda}^{\mu}_{a,j}\right
)\,{\tilde{\bm{x}}^{\mu}}_{a}\right\}\;,\end{split} \tag{85}
$$
where
$$
\tilde{\lambda}_{a,k}=\frac{\bm{\tilde{x}}_{a}\cdot\bm{w}_{k}}{\sqrt{N}}=\sqrt
{1-\Delta}\lambda_{1,k}+\sqrt{\Delta}\lambda_{a+1,k}\,. \tag{86}
$$
While the equations for $b$ and $M$ remain unchanged, we need to include additional terms in the equation for $Q$ . We find
$$
\displaystyle\begin{split}Q^{\mu+1}_{jk}&=Q^{\mu}_{jk}+\frac{\eta}{N}\mathbb{E
}_{\bm{c}}\left\{\left(\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}_{j}
\Lambda_{k}\right]+\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}_{k}\Lambda_
{j}\right]\right)\left[2+\frac{\eta}{B}\left(\sqrt{1-\Delta}\sigma_{1,c_{1}}^{
2}-b((1-\Delta)\sigma_{1,c_{1}}^{2}+\Delta))\right)\right]\right.\\
&+\frac{\eta}{B}\mathbb{E}_{\bm{x}|\bm{c}}\left[\Lambda_{j}\Lambda_{k}\right](
(1-\Delta)\sigma_{1,c_{1}}^{2}+\Delta)\\
&+\frac{\eta}{B}\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}_{j}\tilde{
\lambda}_{k}\right]\left(\sigma_{1,c_{1}}^{2}-2b\sqrt{1-\Delta}\,\sigma_{1,c_{
1}}^{2}+b^{2}((1-\Delta)\sigma_{1,c_{1}}^{2}+\Delta))\right)\\
&+\frac{\eta(B-1)}{B}(1-\Delta)\mathbb{E}_{\bm{x}|\bm{c}}\left[\Lambda_{a,j}
\Lambda_{a^{\prime},k}\right]\sigma_{1,c_{1}}^{2}\\
&+\frac{\eta(B-1)}{B}\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}_{a,j}
\tilde{\lambda}_{a^{\prime},k}\right]\left(\left(1+b^{2}(1-\Delta)\right)
\sigma_{1,c_{1}}^{2}-2b\sqrt{1-\Delta}\sigma_{1,c_{1}}^{2}\right)\\
&+\left.\frac{\eta(B-1)}{B}\left(\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{
\lambda}_{a,j}\Lambda_{a^{\prime},k}\right]+\mathbb{E}_{\bm{x}|\bm{c}}\left[
\tilde{\lambda}_{a,k}\Lambda_{a^{\prime},j}\right]\right)\left(\sqrt{1-\Delta}
\sigma_{1,c_{1}}^{2}-b(1-\Delta)\sigma_{1,c_{1}}^{2}\right)\right\}\end{split} \tag{87}
$$
We derive the following expressions for the average quantities, valid for $a\neq a^{\prime}$
$$
\displaystyle\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}_{a,j}\tilde{
\lambda}_{a^{\prime},k}\right] \displaystyle=(1-\Delta)\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{1,j}\lambda_{
1,k}\right]\,, \displaystyle\mathbb{E}_{\bm{x}|\bm{c}}\left[\tilde{\lambda}_{a,j}\Lambda_{a^{
\prime},k}\right] \displaystyle=\left[\sqrt{1-\Delta}-b(1-\Delta)\right]\mathbb{E}_{\bm{x}|\bm{c
}}\left[\lambda_{1,j}\lambda_{1,k}\right]-(1-\Delta)\sum_{i=1}^{K}Q_{ki}
\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{1,j}\lambda_{1,i}\right]\,, \displaystyle\mathbb{E}_{\bm{x}|\bm{c}}\left[\Lambda_{a,j}\Lambda_{a^{\prime},
k}\right] \displaystyle=(1-b\sqrt{1-\Delta})^{2}\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_
{1,j}\lambda_{1,k}\right]+(1-\Delta)\sum_{i,h=1}^{K}Q_{ji}Q_{kh}\mathbb{E}_{
\bm{x}|\bm{c}}\left[\lambda_{1,i}\lambda_{1,h}\right] \displaystyle+ \displaystyle\left[b(1-\Delta)-\sqrt{1-\Delta}\right]\sum_{i=1}^{K}\left(Q_{ji
}\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{1,k}\lambda_{1,i}\right]+Q_{ki}
\mathbb{E}_{\bm{x}|\bm{c}}\left[\lambda_{1,j}\lambda_{1,i}\right]\right)\,, \tag{88}
$$
where $\Lambda_{a,j}$ is defined as in Eq. (82).
## Appendix B Supplementary figures and additional details
<details>
<summary>x10.png Details</summary>

### Visual Description
## Line Charts: Training Dynamics of Curriculum Learning Strategies
### Overview
The image contains three line charts labeled a), b), and c), arranged with a) and b) in the top row and c) centered below them. The charts collectively visualize the training dynamics of four different learning strategies over a common x-axis, "Training time α". The strategies are distinguished by color and marker style, as defined in a shared legend. The plots track three different metrics: generalization error, cosine similarity with a signal, and the norm of irrelevant weights.
### Components/Axes
* **Common X-Axis (All Plots):** "Training time α". The scale is linear, ranging from 0 to 12 with major tick marks at intervals of 2 (0, 2, 4, 6, 8, 10, 12).
* **Legend (Present in plot a, top-right corner):**
* **Curriculum:** Blue dashed line with circular markers (`--o`).
* **Anti-Curriculum:** Orange dashed line with square markers (`--s`).
* **Optimal (Î):** Black solid line with diamond markers (`-d`).
* **Optimal (Πand η):** Green solid line with 'x' markers (`-x`).
* **Plot a) Y-Axis:** "Generalization error". The scale is logarithmic, with labeled ticks at `2 x 10^-1`, `3 x 10^-1`, and `4 x 10^-1`.
* **Plot b) Y-Axis:** "Cosine similarity with signal". The scale is linear, ranging from 0.0 to 1.0 with ticks at 0.2 intervals.
* **Inset Plot (within plot b, bottom-right):** A zoomed-in view of the later training period (α from 8.0 to 12.0). Its y-axis ranges from 0.89 to 0.97.
* **Plot c) Y-Axis:** "Norm of irrelevant weights". The scale is linear, ranging from 0.5 to 4.0 with ticks at 0.5 intervals.
### Detailed Analysis
**Plot a) Generalization Error vs. Training Time**
* **Trend Verification:** All four lines show a downward trend, indicating decreasing error over training time. The rate of decrease varies significantly.
* **Data Series Analysis:**
* **Optimal (Πand η) [Green, 'x']:** Shows the steepest and most consistent decline. Starts near 4.5 x 10^-1 at α=0 and ends at the lowest point, approximately 1.0 x 10^-1 at α=12.
* **Optimal (Î) [Black, diamond]:** Follows a similar but slightly less steep path than the green line. Ends at approximately 1.5 x 10^-1 at α=12.
* **Curriculum [Blue, circle]:** Declines steadily until about α=6, after which the rate of improvement slows considerably. Ends at approximately 1.7 x 10^-1 at α=12.
* **Anti-Curriculum [Orange, square]:** Shows the slowest rate of decrease. Remains the highest error curve throughout. Ends at approximately 1.8 x 10^-1 at α=12.
**Plot b) Cosine Similarity with Signal vs. Training Time**
* **Trend Verification:** All lines show an upward trend, indicating increasing alignment with the target signal over time. They rise sharply initially and then plateau.
* **Data Series Analysis:**
* **Optimal (Πand η) [Green, 'x']:** Rises the fastest and achieves the highest similarity, approaching ~0.97 by α=12.
* **Optimal (Î) [Black, diamond]:** Very closely follows the green line, ending just slightly below it at α=12.
* **Curriculum [Blue, circle]:** Rises more slowly than the optimal methods. The inset shows it has the lowest similarity in the late stage, ending at ~0.93 at α=12.
* **Anti-Curriculum [Orange, square]:** Initially lags significantly behind the others but shows a steady increase. The inset reveals it surpasses the Curriculum (blue) line around α=9.5, ending at ~0.95 at α=12.
* **Inset Detail:** The inset (x: 8.0-12.0, y: 0.89-0.97) clearly shows the late-stage ordering: Green (highest) > Black > Orange > Blue (lowest).
**Plot c) Norm of Irrelevant Weights vs. Training Time**
* **Trend Verification:** The trends diverge dramatically. The orange and black lines increase, the blue line increases after a delay, and the green line decreases.
* **Data Series Analysis:**
* **Anti-Curriculum [Orange, square]:** Increases rapidly from α=0, plateaus around α=5, and remains constant at the highest value of ~4.2.
* **Optimal (Î) [Black, diamond]:** Begins increasing later (around α=2), rises steadily, and plateaus around α=8 at a value of ~2.8.
* **Curriculum [Blue, circle]:** Remains flat at ~1.0 until α=6, then begins a steady increase, reaching ~2.4 at α=12.
* **Optimal (Πand η) [Green, 'x']:** The only strategy that reduces this norm. It starts at ~1.0 and decreases steadily throughout training, ending at ~0.6 at α=12.
### Key Observations
1. **Performance Hierarchy:** The "Optimal (Πand η)" strategy (green) consistently performs best across all metrics: lowest generalization error, highest signal similarity, and lowest norm of irrelevant weights.
2. **Anti-Curriculum Drawback:** The "Anti-Curriculum" strategy (orange) results in the highest generalization error and causes the norm of irrelevant weights to grow the largest and fastest, suggesting it may be fitting noise.
3. **Delayed Effects:** The "Curriculum" strategy (blue) shows a delayed increase in the norm of irrelevant weights (starting at α=6), which coincides with the slowing of its improvement in generalization error (plot a).
4. **Similarity vs. Generalization:** While all methods achieve high cosine similarity (>0.9) by the end, their generalization errors differ significantly. This indicates that high similarity with the signal alone does not guarantee optimal generalization; controlling irrelevant weights (as the green line does) appears crucial.
### Interpretation
The data demonstrates the comparative effectiveness of different curriculum learning schedules. The "Optimal" strategies, which presumably use a theoretically derived schedule, outperform the simpler Curriculum and Anti-Curriculum approaches.
The key insight is the relationship between the three metrics. Plot c) reveals the mechanism behind the performance differences in plots a) and b). The "Optimal (Πand η)" method is unique in its ability to *suppress* the norm of irrelevant weights while simultaneously increasing similarity with the relevant signal. This selective pressure likely explains its superior generalization. In contrast, the Anti-Curriculum method aggressively increases the norm of irrelevant weights, which correlates with its poor generalization despite achieving reasonable signal similarity. The standard Curriculum method shows a middle ground, where a late-stage increase in irrelevant weight norm may be limiting its final performance.
Therefore, the charts suggest that an effective training strategy must not only align with the target signal but also actively discourage the model from relying on irrelevant features. The "Optimal (Πand η)" schedule appears to successfully balance these two objectives.
</details>
Figure 10: Dynamics of the curriculum learning problem under different training schedulesâcurriculum (easy to hard) at $\eta=3$ , anti-curriculum (hard to easy) at $\eta=3$ , the optimal difficulty protocol at $\eta=3$ (see Fig. 2 b), and the optimal protocol obtained by jointly optimizing $\Delta$ and $\eta$ (see Fig. 3 a). (a) Generalization error vs. normalized training time $\alpha=\mu/N$ . (b) Cosine similarity $M_{11}/\sqrt{TQ_{11}}$ with the target signal (inset zooms into the late-training regime). (c) Squared norm of irrelevant weights $Q_{22}$ vs. $\alpha$ . Parameters: $\alpha_{F}=12$ , $\Delta_{1}=0$ , $\Delta_{2}=2$ , $\eta=3$ , $\lambda=0$ , $T=2$ . Initial conditions: $Q_{11}=Q_{22}=1$ , $M_{11}=0$ .
The initial conditions for the order parameters used in Figs. 7 and 8 are
$$
\displaystyle R=\frac{{\bm{w}}^{\top}{\bm{\mu}}_{\bm{c}}}{N}=\begin{pmatrix}0.
116&0.029\\
-0.005&0.104\end{pmatrix}\,,\qquad Q=\frac{{\bm{w}}^{\top}{\bm{w}}}{N}=\begin{
pmatrix}0.25&0.003\\
0.003&0.25\end{pmatrix}\,, \displaystyle\Omega_{(1,1)(1,1)}=\frac{{\bm{\mu}}_{1,1}\cdot{\bm{\mu}}_{1,1}}{
N}=0.947\,,\qquad\Omega_{(1,2)(1,2)}=\frac{{\bm{\mu}}_{1,2}\cdot{\bm{\mu}}_{1,
2}}{N}=0.990\,. \tag{91}
$$
The initial conditions for the order parameters used in Fig. 9 are
$$
\displaystyle R=\frac{{\bm{w}}^{\top}{\bm{\mu}}_{\bm{c}}}{N}=\begin{pmatrix}0.
339&0.200\\
0.173&0.263\end{pmatrix}\,,\qquad Q=\frac{{\bm{w}}^{\top}{\bm{w}}}{N}=\begin{
pmatrix}1&0.00068\\
0.00068&1\end{pmatrix}\,, \displaystyle\Omega_{(1,1)(1,1)}=\frac{{\bm{\mu}}_{1,1}\cdot{\bm{\mu}}_{1,1}}{
N}=1.737\,,\qquad\Omega_{(1,2)(1,2)}=\frac{{\bm{\mu}}_{1,2}\cdot{\bm{\mu}}_{1,
2}}{N}=1.158\,. \tag{92}
$$
The test set used in Fig. 9 b contains $13996$ examples. The standard deviations of the clusters are $\sigma_{1,1}=0.05$ and $\sigma_{1,2}=0.033$ . The cluster membership probability is $p_{c}([c_{1}=1,c_{2}=1])=0.47$ and $p_{c}([c_{1}=2,c_{2}=1])=0.53$ . The initial conditions for the order parameters used in Fig. 13 are
$$
\displaystyle R=\frac{{\bm{w}}^{\top}{\bm{\mu}}_{\bm{c}}}{N}=\begin{pmatrix}0.
099&-0.005\\
-0.002&0.102\end{pmatrix}\,,\qquad Q=\frac{{\bm{w}}^{\top}{\bm{w}}}{N}=\begin{
pmatrix}0.25&-0.002\\
-0.002&0.25\end{pmatrix}\,, \displaystyle\Omega_{(1,1)(1,1)}=\frac{{\bm{\mu}}_{1,1}\cdot{\bm{\mu}}_{1,1}}{
N}=0.976\,,\qquad\Omega_{(1,2)(1,2)}=\frac{{\bm{\mu}}_{1,2}\cdot{\bm{\mu}}_{1,
2}}{N}=1.014\,. \tag{93}
$$
## Appendix C Numerical simulations
In this appendix, we validate our theoretical predictions against numerical simulations for the three scenarios studied: curriculum learning (Fig. 11), dropout regularization (Fig. 12), and denoising autoencoders (Fig. 13). For each case, the theoretical curves are obtained by numerically integrating the respective ODEs, obtained in the high-dimensional limit $N\to\infty$ . The simulations are instead obtained for a single SGD trajectory at large but finite $N$ . We observe good agreement between theory and simulations.
<details>
<summary>x11.png Details</summary>

### Visual Description
## Multi-Panel Line Plot: Training Dynamics Comparison
### Overview
The image is a 2x2 grid of four line plots, labeled a), b), c), and d). Each plot compares a theoretical prediction (solid blue line) against simulation results (red 'x' markers) for a different metric as a function of "Training time α". The plots collectively illustrate the evolution of various system parameters during a learning or training process.
### Components/Axes
* **Common X-Axis (All Plots):** "Training time α". The scale is linear, ranging from 0 to 5, with major tick marks at integer intervals (0, 1, 2, 3, 4, 5).
* **Plot-Specific Y-Axes:**
* **a) Top-Left:** "Generalization error". Scale is linear, ranging from approximately 0.25 to 0.50.
* **b) Top-Right:** "Mââ". Scale is linear, ranging from 0.0 to just above 2.0.
* **c) Bottom-Left:** "Qââ". Scale is linear, ranging from 1 to just above 6.
* **d) Bottom-Right:** "Qââ". Scale is linear, ranging from 1.0 to 3.5.
* **Legend:** Located in the top-right corner of plot a). It defines the two data series present in all four subplots:
* **Blue Line:** "Theory"
* **Red 'x' Markers:** "Simulations"
### Detailed Analysis
**Plot a) Generalization error vs. Training time α**
* **Trend:** The generalization error shows a clear, monotonic **decreasing trend**. The curve is convex, with the rate of decrease slowing as α increases.
* **Data Points:** The error starts at approximately 0.50 at α=0. It falls to about 0.38 at α=1, 0.32 at α=2, 0.28 at α=3, 0.25 at α=4, and ends near 0.23 at α=5.
* **Agreement:** The simulation points (red 'x') align almost perfectly with the theoretical curve (blue line) across the entire range.
**Plot b) Mââ vs. Training time α**
* **Trend:** The variable Mââ shows a **monotonically increasing trend**. The curve is slightly concave, indicating a gradually slowing rate of increase.
* **Data Points:** Mââ starts at 0.0 at α=0. It rises to approximately 0.5 at α=1, 1.0 at α=2, 1.4 at α=3, 1.8 at α=4, and reaches about 2.2 at α=5.
* **Agreement:** Excellent agreement between theory and simulation is observed throughout.
**Plot c) Qââ vs. Training time α**
* **Trend:** Qââ exhibits a **monotonically increasing trend** with a distinct change in slope. The increase is nearly linear from α=0 to αâ2.5, after which the slope decreases, though the trend remains upward.
* **Data Points:** Qââ starts at 1.0 at α=0. It increases to about 2.5 at α=1, 4.0 at α=2, 5.0 at α=3, 5.8 at α=4, and ends near 6.5 at α=5.
* **Agreement:** The simulation data points closely track the theoretical line, including the subtle change in curvature around α=2.5.
**Plot d) Qââ vs. Training time α**
* **Trend:** Qââ shows a **rapid initial increase followed by saturation**. The curve rises steeply from α=0 to αâ2.5, after which it plateaus, showing almost no further increase.
* **Data Points:** Qââ starts at 1.0 at α=0. It increases sharply to about 2.5 at α=1, 3.2 at α=2, and reaches a plateau value of approximately 3.6 at α=2.5. From α=2.5 to α=5, the value remains constant at ~3.6.
* **Agreement:** The theory and simulations show perfect agreement, with the simulation points lying exactly on the theoretical curve, including the sharp transition to the plateau.
### Key Observations
1. **High-Fidelity Model:** The most striking feature is the near-perfect correspondence between the theoretical predictions (blue lines) and the numerical simulation results (red 'x' markers) across all four metrics and the entire training time range. This indicates the theoretical model is highly accurate.
2. **Diverse Dynamical Behaviors:** The four metrics display fundamentally different learning dynamics:
* **a)** Exponential-like decay (Generalization error).
* **b)** Smooth, saturating growth (Mââ).
* **c)** Piecewise-linear growth with a kink (Qââ).
* **d)** Growth with a sharp saturation point (Qââ).
3. **Critical Transition Point:** A notable event occurs around **α â 2.5**. At this point, the growth of Qââ slows (plot c), and the growth of Qââ halts completely (plot d). This suggests a phase transition or the completion of a specific learning stage at this training time.
### Interpretation
This figure likely comes from a theoretical study of machine learning dynamics, possibly analyzing a linear model or a neural network in a specific regime (e.g., the "lazy" or "rich" learning regime). The parameters Mââ, Qââ, and Qââ are probably components of weight or gradient covariance matrices, which are central to such theoretical analyses.
The data demonstrates that the derived theoretical equations (blue lines) successfully capture the complex, non-trivial temporal evolution of the system observed in simulations. The perfect match validates the theoretical framework.
The distinct behaviors of Qââ and Qââ are particularly insightful. They suggest that different components of the system's internal representation (e.g., different directions in weight space) learn at different rates and may "freeze" at different times. The saturation of Qââ at αâ2.5 while Qââ continues to grow implies that learning becomes constrained to a specific subspace after this point. The continued decrease in generalization error (plot a) after this saturation indicates that refinement within this active subspace still improves performance.
In summary, the figure provides strong evidence for a theoretical model that accurately predicts multi-faceted learning dynamics, revealing a structured, multi-stage training process where different internal parameters evolve according to distinct, well-defined rules.
</details>
Figure 11: Comparison between theory and simulations in the curriculum learning problem: a) generalization error, b) teacher-student overlap $M_{11}$ , c) squared norm $Q_{11}$ of the relevant weights, and d) squared norm $Q_{22}$ of the irrelevant weights. The continuous blue lines have been obtained by integrating numerically the ODEs in Eqs. (48), while the red crosses are the results of numerical simulations of a single trajectory with $N=30000$ . The protocol is anti-curriculum with equal proportion of easy and hard samples. Parameters: $\alpha_{F}=5$ , $\lambda=0$ , $\eta=3$ , $\Delta_{1}=0$ , $\Delta_{2}=2$ , $T_{11}=1$ . Initial conditions: $Q_{11}=0.984$ , $Q_{22}=0.998$ , $M_{11}=0.01$ .
<details>
<summary>x12.png Details</summary>

### Visual Description
## Line Plots: Comparison of Theory vs. Simulations for Learning Metrics
### Overview
The image displays a 2x2 grid of four line plots, labeled a), b), c), and d). Each plot compares a theoretical prediction (solid blue line) against simulation results (red 'x' markers) for a different metric as a function of "Training time α". The plots demonstrate a very close agreement between theory and simulation across all four metrics. A single legend, located in the top-left corner of subplot a), defines the two data series.
### Components/Axes
* **Common X-Axis (All Plots):** Label: "Training time α". Scale: Linear, ranging from 0 to 5 with major tick marks at integer intervals (0, 1, 2, 3, 4, 5).
* **Legend:** Positioned in the top-left corner of subplot a). Contains two entries:
* A solid blue line labeled "Theory".
* A red 'x' marker labeled "Simulations".
* **Subplot a) - Top Left:**
* **Y-Axis Label:** "Generalization error".
* **Y-Axis Scale:** Linear, ranging from approximately 0.04 to 0.17.
* **Subplot b) - Top Right:**
* **Y-Axis Label:** "Mâ,â" (M with subscript 1,1).
* **Y-Axis Scale:** Linear, ranging from 0.0 to approximately 0.45.
* **Subplot c) - Bottom Left:**
* **Y-Axis Label:** "Qââ" (Q with subscript 11).
* **Y-Axis Scale:** Linear, ranging from 0.00 to approximately 0.33.
* **Subplot d) - Bottom Right:**
* **Y-Axis Label:** "Qââ" (Q with subscript 22).
* **Y-Axis Scale:** Linear, ranging from 0.00 to approximately 0.33.
### Detailed Analysis
**Trend Verification & Data Points (Approximate):**
* **Subplot a) Generalization Error:**
* **Trend:** The line slopes downward, showing a decreasing, convex curve. The rate of decrease slows as α increases.
* **Data Points (α, Value):**
* α=0: ~0.165
* α=1: ~0.105
* α=2: ~0.070
* α=3: ~0.050
* α=4: ~0.042
* α=5: ~0.038
* **Subplot b) Mâ,â:**
* **Trend:** The line slopes upward, showing an increasing, concave curve. The rate of increase slows as α increases.
* **Data Points (α, Value):**
* α=0: 0.00
* α=1: ~0.20
* α=2: ~0.32
* α=3: ~0.38
* α=4: ~0.42
* α=5: ~0.44
* **Subplot c) Qââ:**
* **Trend:** The line slopes upward, showing an increasing, concave curve. The rate of increase slows as α increases.
* **Data Points (α, Value):**
* α=0: 0.00
* α=1: ~0.12
* α=2: ~0.21
* α=3: ~0.26
* α=4: ~0.30
* α=5: ~0.32
* **Subplot d) Qââ:**
* **Trend:** The line slopes upward, showing an increasing, concave curve. The rate of increase slows as α increases. The curve and final value are very similar to Qââ.
* **Data Points (α, Value):**
* α=0: 0.00
* α=1: ~0.13
* α=2: ~0.22
* α=3: ~0.27
* α=4: ~0.30
* α=5: ~0.32
**Component Isolation & Cross-Reference:**
* In all four subplots, the red 'x' markers (Simulations) lie directly on or extremely close to the solid blue line (Theory), confirming excellent agreement.
* The legend in subplot a) applies to all four panels, as the same marker and line styles are used throughout.
### Key Observations
1. **Strong Theory-Simulation Agreement:** The most prominent feature is the near-perfect overlap between the theoretical curves and the simulation data points across all metrics and all values of α.
2. **Consistent Functional Forms:** Three of the four metrics (Mâ,â, Qââ, Qââ) show a similar increasing, saturating trend. Generalization error shows the inverseâa decreasing, saturating trend.
3. **Asymptotic Behavior:** All metrics appear to approach an asymptotic limit as training time α increases beyond 4 or 5.
4. **Similarity of Qââ and Qââ:** The plots for Qââ and Qââ are nearly identical in shape and magnitude, suggesting these two quantities evolve similarly during training.
### Interpretation
This figure validates a theoretical model for the dynamics of a learning system. The "Training time α" likely represents a normalized measure of training duration or sample size.
* **What the data suggests:** The theory accurately predicts how key system parameters (Mâ,â, Qââ, Qââ) evolve and how the generalization error decreases during the learning process. The close match implies the theoretical assumptions and derivations are sound for the simulated scenario.
* **Relationship between elements:** The decreasing generalization error (plot a) is causally linked to the increasing values of the internal parameters M and Q (plots b-d). As the system's internal state (represented by M and Q) evolves away from its initial condition (zero), its ability to generalize to new data improves (error drops).
* **Notable patterns:** The saturating trends indicate a learning process that experiences diminishing returns; initial training yields rapid improvement, but gains slow as the system approaches a stable, trained state. The near-identity of Qââ and Qââ might indicate symmetry in the system's structure or that these parameters represent similar properties (e.g., variances of different but equivalent components).
* **Underlying context:** This type of analysis is common in statistical mechanics of learning, mean-field theory of neural networks, or the study of online learning algorithms, where theoretical curves are derived from analytical models and tested against numerical simulations of the learning dynamics. The variables M and Q are typical notations for order parameters in such theories.
</details>
Figure 12: Comparison between theory and simulations for dropout regularization: a) generalization error, b) teacher-student overlap $M_{1,1}$ , c) squared norm $Q_{11}$ , and d) squared norm $Q_{22}$ . The continuous blue lines have been obtained by integrating numerically the ODEs in Eqs. (52)-(53), while the red crosses are the results of numerical simulations of a single trajectory with $N=30000$ . Parameters: $\alpha_{F}=5$ , $\eta=1$ , $\sigma_{n}=0.3$ , $p(\alpha)=p_{f}=0.7$ , $T_{11}=1$ . Initial conditions: $Q_{ij}=M_{nk}=0$ .
<details>
<summary>x13.png Details</summary>

### Visual Description
## Multi-Panel Line Chart: Theory vs. Simulations Over Training Time
### Overview
The image displays a 2x2 grid of four line charts, labeled a), b), c), and d). Each chart compares a theoretical prediction (solid blue line) against simulation results (red 'x' markers) for a different metric plotted against a common x-axis, "Training time α". The overall purpose is to validate a theoretical model by showing its close agreement with empirical simulation data across multiple quantities.
### Components/Axes
* **Common X-Axis (All Plots):** Label: "Training time α". Scale: Linear, ranging from 0 to 5, with major ticks at integer intervals (0, 1, 2, 3, 4, 5).
* **Legend (Located in top-left of plot a):**
* Blue Line: "Theory"
* Red 'x' Marker: "Simulations"
* **Plot a) (Top-Left):**
* **Y-Axis Label:** "MSE(α)-MSE(0)"
* **Y-Axis Scale:** Linear, ranging from approximately -0.175 to 0.000.
* **Plot b) (Top-Right):**
* **Y-Axis Label:** "R_{1,(1,1)}"
* **Y-Axis Scale:** Linear, ranging from 0.10 to 0.45.
* **Plot c) (Bottom-Left):**
* **Y-Axis Label:** "Q_{1,1}"
* **Y-Axis Scale:** Linear, ranging from 0.10 to 0.24.
* **Plot d) (Bottom-Right):**
* **Y-Axis Label:** "Q_{2,2}"
* **Y-Axis Scale:** Linear, ranging from 0.10 to 0.24.
### Detailed Analysis
**Trend Verification & Data Points:**
* **Plot a) MSE(α)-MSE(0):**
* **Trend:** The curve slopes downward monotonically, starting at 0 and decreasing at a decelerating rate.
* **Data Points (Approximate):**
* α=0: y=0.000
* α=1: yâ -0.090
* α=2: yâ -0.130
* α=3: yâ -0.160
* α=4: yâ -0.170
* α=5: yâ -0.175
* **Agreement:** Simulation markers (red 'x') lie almost perfectly on the theoretical blue line throughout.
* **Plot b) R_{1,(1,1)}:**
* **Trend:** The curve slopes upward in a sigmoidal (S-shaped) fashion, starting slowly, accelerating in the middle, and then leveling off.
* **Data Points (Approximate):**
* α=0: yâ 0.10
* α=1: yâ 0.12
* α=2: yâ 0.22
* α=3: yâ 0.35
* α=4: yâ 0.42
* α=5: yâ 0.44
* **Agreement:** Excellent match between simulations and theory across the entire range.
* **Plot c) Q_{1,1}:**
* **Trend:** The curve is U-shaped (convex). It decreases to a minimum and then increases.
* **Data Points (Approximate):**
* α=0: yâ 0.25
* α=1: yâ 0.11 (near minimum)
* α=2: yâ 0.12
* α=3: yâ 0.18
* α=4: yâ 0.22
* α=5: yâ 0.24
* **Agreement:** Very strong alignment. The simulation points trace the theoretical curve precisely, including the minimum around 뱉1.5.
* **Plot d) Q_{2,2}:**
* **Trend:** Nearly identical U-shaped (convex) trend to plot c).
* **Data Points (Approximate):**
* α=0: yâ 0.25
* α=1: yâ 0.11 (near minimum)
* α=2: yâ 0.12
* α=3: yâ 0.18
* α=4: yâ 0.22
* α=5: yâ 0.24
* **Agreement:** Again, near-perfect correspondence between the red simulation markers and the blue theoretical line.
### Key Observations
1. **High-Fidelity Validation:** The most striking observation is the exceptional agreement between the theoretical model (blue line) and the simulation results (red 'x's) across all four metrics and the entire range of training time α. The simulation points show minimal scatter around the theoretical prediction.
2. **Diverse Metric Behaviors:** The four tracked quantities exhibit fundamentally different temporal dynamics:
* **MSE Difference (a):** Monotonic decrease.
* **R Metric (b):** Monotonic, sigmoidal increase.
* **Q Metrics (c, d):** Non-monotonic, U-shaped behavior with a clear minimum.
3. **Similarity of Q Metrics:** The plots for Q_{1,1} and Q_{2,2} are visually almost identical, suggesting these two components of the system evolve in a very similar manner over training time.
### Interpretation
This set of charts serves as a strong empirical validation of a theoretical framework describing a learning or optimization process (indicated by "Training time α"). The near-perfect overlap between theory and simulation suggests the theoretical model successfully captures the essential dynamics of the system.
The different trends reveal the complex, multi-faceted nature of the process:
* The decreasing **MSE(α)-MSE(0)** in plot a) indicates the system's error (or loss) is consistently improving relative to its initial state as training progresses.
* The increasing **R_{1,(1,1)}** in plot b) suggests a correlation or response metric is strengthening over time.
* The U-shaped **Q_{1,1}** and **Q_{2,2}** in plots c) and d) are particularly insightful. They imply an **optimal training duration** (around α â 1.5) where these specific quantities are minimized. Training for shorter or longer than this optimal point results in higher values for Q. This could represent a trade-off, such as the balance between learning speed and stability, or the evolution of internal model parameters.
In summary, the data demonstrates that the theoretical model is highly accurate, and it reveals that the learning process involves a combination of monotonic improvements (in error and correlation) alongside non-monotonic, optimized evolution of internal system states (the Q metrics). The consistency between Q_{1,1} and Q_{2,2} may indicate symmetry or similar roles for these components within the modeled system.
</details>
Figure 13: Comparison between theory and simulations for the denoising autoencoder model: a) mean square error improvement, b) student-centroid overlap $R_{1,(1,1)}$ , c) squared norm $Q_{11}$ . The continuous blue lines have been obtained by integrating numerically the ODEs in Eqs. (80) and (87), while the red crosses are the results of numerical simulations of a single trajectory with $N=10000$ . Parameters: $\alpha_{F}=1$ , $\eta=2$ , $B(\alpha)=\bar{B}=5$ , $K=C_{1}=2$ , $\sigma=0.1$ , $g(z)=z$ . The skip connection $b$ is fixed ( $\eta_{b}=0$ ) to the optimal value in Eq. (26). Initial conditions are given in Eq. (93).