shape bias

**Shape Bias** is the **reliance on global shape features (contours, silhouettes, structural geometry) for object recognition** — shape-biased models, like human visual perception, classify objects primarily by their shape rather than texture, leading to more robust and human-aligned representations. **Inducing Shape Bias** - **Stylization**: Train on style-transferred images (random textures applied to ImageNet images) — forces the model to ignore texture. - **Data Augmentation**: Use augmentations that preserve shape but alter texture (color jittering, style transfer, texture randomization). - **Architecture**: Vision Transformers (ViTs) naturally exhibit more shape bias than CNNs due to their global attention mechanism. - **Multi-Crop**: Random cropping at different scales encourages attending to global structure. **Why It Matters** - **Robustness**: Shape-biased models are more robust to distribution shifts, noise, and adversarial perturbations. - **Transfer**: Shape features transfer better to new domains than texture features. - **Human Alignment**: Shape bias aligns model representations with human visual processing — better interpretability. **Shape Bias** is **seeing the forest, not just the trees** — prioritizing global shape over local texture for robust, human-aligned visual recognition.