Neural Circuits & Behavior Progress Report // Oh Hyeon Choung - Recurrent Networks but not Classic CNNs Explain Global Shape Processing

Feedforward Convolutional Neural Networks (ffCNNs) have become state-of-the-art models in computer vision and neuroscience. However, ffCNNs only roughly mimic human vision. They lack recurrent connections and rely mainly on local features, contrary to humans who use global shape computations. Here, using visual crowding as a well-controlled and specific probe for global shape computations, we show that ffCNNs cannot produce human-like processing for principled architectural reasons. We provide evidence that recurrent processing is crucial for global shape computations in both humans and machines. In particular, we show that Capsule Networks, combining ffCNNs with a time-consuming recurrent grouping and segmentation mechanism, can solve this challenge in a natural manner. Together, we suggest recurrent grouping and segmentation processes are essential to understand the human visual system and create better models that harness global shape computations.