## Conceptual Diagram: Few-shot vs. Zero-shot Learning Representations ### Overview This image presents two side-by-side conceptual diagrams illustrating the difference between "Few-shot" and "Zero-shot" learning paradigms within a shared feature space. The space is partitioned into three distinct regions (colored green, orange, and blue), representing different classes. The diagrams contrast how class prototypes (centroids) are established: via observed data samples in Few-shot learning, and via semantic vectors in Zero-shot learning. ### Components/Axes * **Spatial Layout:** Both diagrams share an identical 2D feature space divided into three Voronoi-like regions: * **Left (Green):** Associated with centroid `c1`. * **Top-Right (Orange):** Associated with centroid `c2`. * **Bottom-Right (Blue):** Associated with centroid `c3`. * **Common Elements:** * **`x` (White Circle):** Located near the center of the diagram, at the intersection of the three regions. This represents a query point to be classified. * **`c1`, `c2`, `c3` (Black Circles):** Represent the class prototypes or centroids. * **Dashed Lines:** Connect the query point `x` to each centroid (`c1`, `c2`, `c3`), indicating distance-based relationships. * **Panel (a) - Few-shot Specifics:** * **Small Colored Circles:** Clusters of data points surrounding each centroid. * **Arrows:** Pointing from the small circles toward the respective centroid, indicating the aggregation of samples to form the prototype. * **Panel (b) - Zero-shot Specifics:** * **`v1`, `v2`, `v3`:** Labels with curved arrows pointing to centroids `c1`, `c2`, and `c3` respectively. These represent semantic vectors or embeddings. ### Detailed Analysis #### Panel (a): Few-shot * **Visual Trend:** The diagram shows three distinct clusters. * **Data Points:** * **`c1` (Green Region):** Surrounded by 5 small green circles. Arrows point from these circles to `c1`. * **`c2` (Orange Region):** Surrounded by 5 small orange circles. Arrows point from these circles to `c2`. * **`c3` (Blue Region):** Surrounded by 5 small blue circles. Arrows point from these circles to `c3`. * **Relationship:** The centroid is visually depicted as the center of mass or representative point derived from the surrounding observed data samples. The query point `x` is positioned such that it is closest to `c1` (based on the visual length of the dashed line), suggesting it would be classified as class 1. #### Panel (b): Zero-shot * **Visual Trend:** The clusters of small data points are absent. Instead, the centroids are defined by external vectors. * **Data Points:** * **`v1`:** An arrow points from the label `v1` to the centroid `c1`. * **`v2`:** An arrow points from the label `v2` to the centroid `c2`. * **`v3`:** An arrow points from the label `v3` to the centroid `c3`. * **Relationship:** The centroids `c1`, `c2`, and `c3` are not derived from observed data points but are projected into the feature space using the semantic vectors `v1`, `v2`, and `v3`. The query point `x` remains in the same relative position, and the classification logic (distance to centroid) remains identical to the Few-shot model. ### Key Observations * **Structural Consistency:** The underlying geometry (the regions and the position of `x`) is identical in both diagrams, emphasizing that the classification *mechanism* (distance-based) is the same, while the *input* to that mechanism differs. * **Source of Truth:** In Few-shot learning, the "truth" is empirical (observed samples). In Zero-shot learning, the "truth" is semantic (vectors/embeddings). * **Absence of Data:** The most notable difference is the complete removal of the small colored circles in the Zero-shot diagram, visually reinforcing the "zero" aspect of the learning paradigm (no training samples). ### Interpretation This diagram effectively visualizes the transition from empirical learning to semantic learning. * **Few-shot Learning:** Demonstrates that the model learns to classify by observing a small number of examples. The centroid is an emergent property of the data distribution. * **Zero-shot Learning:** Demonstrates that the model classifies by mapping semantic concepts (like word embeddings or attributes) into the same feature space as the query point. The centroid is a projected property of the semantic vector. * **Why it matters:** This comparison highlights that both methods rely on a shared feature space. The diagram suggests that if a model can learn to align semantic vectors (`v`) with the feature space, it can perform classification without ever having seen a single training example for that specific class, provided the semantic vector is sufficiently descriptive.

## Diagram: Few-shot vs Zero-shot Classification Models ### Overview The image presents two comparative diagrams illustrating conceptual differences between few-shot and zero-shot learning approaches. Both diagrams use geometric shapes (triangles) divided into colored regions (green, orange, blue) to represent different components or categories. Central nodes (C1, C2, C3) and vectors (V1, V2, V3) are connected through directional arrows, suggesting relationships between elements. ### Components/Axes **Diagram (a) Few-shot:** - **Regions:** - Green triangle (left): Contains C1 with 4 green nodes connected via dotted lines - Orange triangle (top): Contains C2 with 3 orange nodes connected via dotted lines - Blue triangle (bottom): Contains C3 with 3 blue nodes connected via dotted lines - **Central Node:** White circle labeled "X" at the intersection of all regions - **Connections:** Dotted lines from X to C1, C2, C3 **Diagram (b) Zero-shot:** - **Regions:** Same color-coded triangles as (a) - **Central Node:** White circle labeled "X" at the intersection - **Vectors:** Solid arrows labeled V1, V2, V3 pointing from X to C1, C2, C3 respectively - **Additional Elements:** Dashed lines connecting X to C1, C2, C3 ### Detailed Analysis **Few-shot Diagram (a):** - C1 (green) has 4 connected nodes (likely training examples) - C2 (orange) has 3 connected nodes - C3 (blue) has 3 connected nodes - All connections to X are dotted, suggesting uncertain or probabilistic relationships **Zero-shot Diagram (b):** - Vectors V1-V3 show directional influence from X to each C - Solid arrows indicate stronger, more defined relationships - Dashed lines maintain conceptual connections from X to all Cs ### Key Observations 1. **Connection Types:** Few-shot uses dotted lines (uncertainty), zero-shot uses solid vectors (certainty) 2. **Node Distribution:** Few-shot shows more nodes per category (C1 has 4 vs C2/C3 with 3) 3. **Spatial Relationships:** Both diagrams share identical regional structure but differ in connection types 4. **Vector Directionality:** Zero-shot explicitly shows directional influence from X to categories ### Interpretation The diagrams illustrate a progression from exploratory (few-shot) to definitive (zero-shot) learning paradigms: - **Few-shot (a):** Represents initial model exploration where relationships between central node X and categories (C1-C3) are uncertain, shown by dotted connections. The varying number of nodes per category suggests asymmetric training data distribution. - **Zero-shot (b):** Demonstrates a matured model where X exerts directional influence (via vectors V1-V3) on all categories, indicating learned relationships. The maintained dashed lines suggest retained conceptual connections despite more defined vector relationships. The color-coded regions likely represent different semantic categories or feature spaces. The transition from dotted to solid connections implies that zero-shot learning resolves uncertainties present in few-shot approaches through learned vector representations. The consistent regional structure across both diagrams suggests these models share foundational architecture but differ in their operational phases or training stages.