Shape Bias

Shape bias in artificial vision systems refers to the extent to which models prioritize shape information over texture when classifying objects, a characteristic that contrasts with human vision's stronger reliance on shape. Current research focuses on understanding how different model architectures, including convolutional neural networks (CNNs) and transformers, exhibit varying degrees of shape bias, and how this bias relates to generalization performance, robustness to adversarial attacks, and alignment with human perception. This research is significant because it helps clarify the limitations of current deep learning models, potentially leading to improved model design and more robust, human-like artificial vision systems.