Search Papers | Poster Sessions | All Posters
Poster A105 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
Configural-Shape Representation in Deep Neural Networks
Fenil R. Doshi1,2 (), Talia Konkle1,2, George A. Alvarez1,2; 1Harvard University, 2Kempner Institute for the Study of Natural and Artificial Intelligence
We introduce a ‘configural shape index’ to quantify the quality of configural shape information in deep neural networks used to model human visual processing. Unlike shape-vs-texture bias measures (Geirhos et al. 2018), which capture the relative importance of shape in making classification decisions, our index captures the quality of shape representations in absolute terms (not relative to texture), and can be applied to any layer of any DNN model, regardless of model objective. Over a set of 92 models (including CNNs and transformers trained on a variety of tasks), we find low to modest sensitivity to configural shape, even in models with near human levels of shape-bias. These results suggest that there remains significant room for improving the quality of configural shape representations in DNN models of object recognition.
Keywords: Shape bias Holistic processing Deep Neural Networks Mid-level vision