## Diagram: Shifts in Machine Learning Robustness ### Overview The image is a diagram illustrating different types of shifts in machine learning data, categorized by whether they are handled with improved robustness from Feature Matching (FMs) or remain persistently challenging. The diagram is divided into two rows, representing In-Distribution (ID) and Out-of-Distribution (OOD) data, and several columns, each representing a different type of shift. ### Components/Axes * **Title:** Shifts with improved robustness from FMs (left side) and Persistently challenging shifts (right side). * **Rows:** * ID (In-Distribution): Top row, light blue background. * OOD (Out-of-Distribution): Bottom row, light pink background. * **Columns (Left Side):** * Common corruptions * Shifts across space * Domain shift * **Columns (Right Side):** * Extrapolation, e.g. shift across time * Spurious correlations * **Arrows:** Downward pointing arrows connect each ID image/text to its corresponding OOD image/text. * **Citations:** Citations are listed below each column, indicating the source of the example. ### Detailed Analysis **Left Side: Shifts with improved robustness from FMs** * **Common corruptions:** * ID: Image of a yellow bird perched on a branch with pink flowers. * OOD: Image of the same bird, but with a blurred or corrupted appearance. * Citation: Hendrycks '19 * **Shifts across space:** * ID: Aerial view of a landscape with fields and roads. * OOD: A pixelated or distorted version of the same landscape. * Citation: Xie '21 * **Domain shift:** * ID: Photograph of a bunch of bananas and some kiwis. * OOD: A black and white sketch of bananas. * Citation: Radford '21 **Right Side: Persistently challenging shifts** * **Extrapolation, e.g. shift across time:** * ID: Text "Pence is the Vice President of the US." * OOD: Text "Harris is the Vice President of the US." * Citation: Lazaridou '21 * **Spurious correlations:** * ID: Image of a cow in a mountainous landscape. * OOD: Image of a cow lying on a beach near a boat. * Citation: Beery '18 ### Key Observations * The left side of the diagram shows shifts that are becoming more manageable due to advancements in Feature Matching (FMs). * The right side highlights shifts that remain difficult for machine learning models to handle. * The shift from ID to OOD represents a change in the data distribution that can negatively impact model performance. ### Interpretation The diagram illustrates the ongoing challenges in machine learning robustness. While models are improving at handling certain types of shifts (e.g., common corruptions, domain shifts), others (e.g., extrapolation, spurious correlations) continue to pose significant problems. The distinction between ID and OOD data is crucial for understanding model generalization and the ability to perform well on unseen data. The examples provided highlight the diverse nature of these shifts, ranging from image corruptions to changes in semantic meaning over time. The diagram suggests that future research should focus on developing methods that can effectively address these persistently challenging shifts to improve the reliability and trustworthiness of machine learning systems.

\n ## Diagram: Robustness to Distribution Shifts in Machine Learning ### Overview This diagram illustrates different types of distribution shifts encountered in machine learning and their impact on model performance. It categorizes these shifts into those where robustness can be improved with Feature Manipulation (FMs) and those that remain persistently challenging. The diagram visually represents the shift from In-Distribution (ID) data to Out-of-Distribution (OOD) data using image pairs and associated research papers. ### Components/Axes The diagram is divided into two main sections: "Shifts with improved robustness from FMs" (left) and "Persistently challenging shifts" (right). Each section contains three examples of distribution shifts, each with an ID image, an OOD image, and a corresponding research paper citation. The diagram uses color-coding: blue for ID data and pink for OOD data. Arrows indicate the shift from ID to OOD. The categories of shifts are: * **Common corruptions** * **Shifts across space** * **Domain shift** * **Extrapolation, e.g. shift across time** * **Spurious correlations** ### Detailed Analysis or Content Details **Section 1: Shifts with improved robustness from FMs** * **Common corruptions:** * ID Image: A yellow bird perched on a branch. * OOD Image: A similar yellow bird, but with a slightly different pose and background. * Citation: Hendrycks '19 * **Shifts across space:** * ID Image: A forest scene with visible trees and foliage. * OOD Image: A desert landscape with sparse vegetation. * Citation: Xie '21 * **Domain shift:** * ID Image: A bunch of bananas. * OOD Image: A metallic, reflective surface resembling a banana peel. * Citation: Radford '21 **Section 2: Persistently challenging shifts** * **Extrapolation, e.g. shift across time:** * ID Image: A portrait of Mike Pence with the text "Pence is the Vice President of the US." * OOD Image: A portrait of Kamala Harris with the text "Harris is the Vice President of the US." * Citation: Lazaridou '21 * **Spurious correlations:** * ID Image: A cow standing in a grassy field. * OOD Image: A cow on a beach with a boat in the background. * Citation: Beery '18 ### Key Observations The diagram highlights that certain types of distribution shifts are more easily addressed with techniques like Feature Manipulation. These include common corruptions, shifts across space, and domain shifts. However, shifts involving extrapolation (like changes over time) and spurious correlations pose more significant challenges. The examples demonstrate how seemingly small changes in the input data can lead to incorrect predictions when the model encounters OOD data. ### Interpretation This diagram illustrates a core challenge in machine learning: the gap between training data (ID) and real-world data (OOD). The categorization of shifts suggests that the nature of the shift significantly impacts the effectiveness of mitigation strategies. Shifts that involve changes in image style or minor variations in the environment can be addressed with techniques that enhance the model's robustness to these changes. However, shifts that require reasoning about time or understanding underlying causal relationships (like spurious correlations) are much harder to handle. The inclusion of research citations indicates ongoing efforts to address these challenges. The example of the Vice President shift is particularly insightful, demonstrating that even a simple change in a key attribute (the person in the image) can lead to a significant performance drop if the model relies on spurious correlations. This diagram is a visual representation of the need for models that can generalize beyond the training distribution and reason about the world in a more robust and reliable way.

## Diagram: Foundation Model Robustness to Distribution Shifts ### Overview This image is a conceptual diagram illustrating two categories of distribution shifts encountered by foundation models (FMs). It contrasts shifts where FMs have shown improved robustness against those that remain persistently challenging. The diagram uses a combination of example images, text, and citations to categorize and explain these shifts. ### Components/Axes The diagram is organized into two main vertical panels, each with a purple header. **Left Panel Header:** "Shifts with improved robustness from FMs" **Right Panel Header:** "Persistently challenging shifts" A vertical axis on the far left defines two states: * **ID** (In-Distribution): Represented by a light blue background. * **OOD** (Out-of-Distribution): Represented by a light pink background. Each panel contains two to three columns representing specific shift types. Each column has: 1. A category title at the top. 2. An example image or text block in the ID (blue) section. 3. A downward-pointing purple arrow. 4. A corresponding example image or text block in the OOD (pink) section. 5. A citation (Author 'Year) at the very bottom in blue text. ### Detailed Analysis #### Left Panel: Shifts with improved robustness from FMs This panel contains three columns: 1. **Column 1: Common corruptions** * **ID Image (Top-Left):** A clear, color photograph of a yellow bird perched on a branch with pink blossoms. * **OOD Image (Bottom-Left):** A blurred, lower-resolution version of the same bird image. * **Citation:** Hendrycks '19 2. **Column 2: Shifts across space** * **ID Image (Top-Center):** An aerial or satellite photograph of a landscape with fields and roads. * **OOD Image (Bottom-Center):** A heavily pixelated or low-resolution version of the same landscape image. * **Citation:** Xie '21 3. **Column 3: Domain shift** * **ID Image (Top-Right):** A photograph of a bunch of yellow bananas in a wooden bowl. * **OOD Image (Bottom-Right):** A photograph of a single, peeled banana against a white background. * **Citation:** Radford '21 #### Right Panel: Persistently challenging shifts This panel contains two columns: 1. **Column 1: Extrapolation, e.g. shift across time** * **ID Text (Top-Left):** "Pence is the Vice President of the US." * **OOD Text (Bottom-Left):** "Harris is the Vice President of the US." * **Citation:** Lazaridou '21 2. **Column 2: Spurious correlations** * **ID Image (Top-Right):** A photograph of a cow standing on a grassy mountain slope. * **OOD Image (Bottom-Right):** A photograph of a cow lying on a sandy beach near a boat. * **Citation:** Beery '18 ### Key Observations * **Visual vs. Semantic Shifts:** The left panel primarily illustrates *visual* corruptions and domain changes (blur, pixelation, object state). The right panel illustrates *semantic* or *contextual* shifts (factual knowledge over time, object-context associations). * **Layout Symmetry:** Both panels use an identical ID (top/blue) to OOD (bottom/pink) flow, connected by arrows, creating a clear comparative structure. * **Citation Placement:** All academic citations are placed at the bottom of their respective columns, attributing the example or the research on that shift type. * **Color Coding:** The light blue (ID) and light pink (OOD) backgrounds are consistently applied across both panels to denote the distribution state. ### Interpretation This diagram serves as a taxonomy for understanding the limitations and strengths of current foundation models regarding distribution shift. It suggests that FMs have become notably robust to many *visual* perturbations and corruptions (left panel), likely due to training on vast, diverse datasets that implicitly cover these variations. Examples like "Common corruptions" (blur) and "Domain shift" (bananas to peeled banana) represent changes in the visual rendering or style of a concept, which models can often generalize across. However, the right panel highlights fundamental challenges that are not merely visual. "Extrapolation across time" involves updating factual knowledge, a task requiring temporal reasoning or access to current information beyond a static training set. "Spurious correlations" involve decoupling objects from their typical backgrounds (e.g., cows are not *only* found on grass), which requires models to learn causal features rather than statistical shortcuts. These shifts are "persistently challenging" because they test deeper reasoning, world knowledge, and the ability to avoid biased associations, pointing to areas where model architecture or training paradigms may need advancement. The diagram effectively argues that robustness is not a monolithic property but is highly dependent on the *nature* of the shift.

## Diagram: Shifts in Robustness and Challenges for Machine Learning Models ### Overview The diagram compares two categories of shifts affecting machine learning (ML) model robustness: 1. **Shifts with improved robustness from FMs** (left section, blue background) 2. **Persistently challenging shifts** (right section, pink background) Each category contains subcategories with visual examples, textual annotations, and citations. The diagram uses spatial grounding (arrows, labels) to illustrate relationships between concepts. --- ### Components/Axes #### Main Sections - **Left Section (Improved Robustness from FMs)** - Subcategories: 1. **Common corruptions** - Image: Yellow bird on a branch (ID) → Blurred bird (OOD) - Citation: Hendrycks '19 2. **Shifts across space** - Image: Forest floor (ID) → Abstract texture (OOD) - Citation: Xie '21 3. **Domain shift** - Image: Bananas (ID) → Banana peel (OOD) - Citation: Radford '21 - **Right Section (Persistently Challenging Shifts)** - Subcategories: 1. **Extrapolation, e.g., shift across time** - Text: "Pence is the Vice President of the US." (ID) → "Harris is the Vice President of the US." (OOD) - Citation: Lazaridou '21 2. **Spurious correlations** - Image: Cow lying in grass (ID) → Cow lying on beach (OOD) - Citation: Beery '18 #### Additional Elements - **ID/OOD Labels**: - **ID** (In-Distribution): Blue text, left side. - **OOD** (Out-Of-Distribution): Red text, right side. - **Arrows**: Connect ID/OOD pairs to their respective subcategories. - **Color Coding**: - Blue: Improved robustness subcategories. - Pink: Persistently challenging subcategories. - Red: OOD labels. --- ### Detailed Analysis #### Left Section (Improved Robustness) 1. **Common corruptions**: - ID: Clear image of a yellow bird on a branch. - OOD: Blurred version of the same bird. - Demonstrates robustness to visual noise. 2. **Shifts across space**: - ID: Natural forest floor texture. - OOD: Abstract, distorted texture. - Tests model generalization to spatial variations. 3. **Domain shift**: - ID: Realistic banana bunch. - OOD: Close-up of a banana peel. - Highlights sensitivity to object-level changes. #### Right Section (Persistently Challenging) 1. **Extrapolation**: - Textual example: Misattribution of U.S. Vice President (Pence → Harris). - Illustrates temporal shifts in factual knowledge. 2. **Spurious correlations**: - ID: Cow lying in grass (natural context). - OOD: Cow lying on a beach (unrelated context). - Shows failure to distinguish contextually irrelevant patterns. --- ### Key Observations - **Visual Contrast**: - ID/OOD pairs use color (blue/pink) to differentiate robustness categories. - Arrows spatially ground relationships between concepts. - **Citations**: - All examples are attributed to specific studies (e.g., Hendrycks '19, Lazaridou '21). - **Textual Anomalies**: - The "Pence/Harris" example is factually incorrect as of 2023 (current VP is Kamala Harris). --- ### Interpretation The diagram emphasizes the importance of addressing **domain shifts**, **temporal extrapolation**, and **spurious correlations** to improve ML robustness. - **Improved robustness** (left) focuses on visual and spatial challenges, while **persistently challenging shifts** (right) highlight conceptual and contextual failures. - The "Pence/Harris" example underscores the risk of models relying on outdated or contextually fragile knowledge. - **OOD** (Out-Of-Distribution) examples (e.g., blurred bird, banana peel) stress the need for models to handle unseen data distributions. This framework aligns with Peircean investigative principles, urging scrutiny of how models generalize across time, space, and context. The diagram advocates for robustness testing against both common and rare failure modes.