October 29, 2020

The relative coding strength of object identity and nonidentity features in human occipito-temporal cortex and convolutional neural networks

Any given visual object input is characterized by multiple visual features, such as identity, position and size. Despite the usefulness of identity and nonidentity features in vision and their joint coding throughout the primate ventral visual processing pathway, they have so far been studied relatively independently. Here we document the relative coding strength of object identity and nonidentity features in a brain region and how this may change across the human ventral visual pathway. We examined a total of four nonidentity features, including two Euclidean features (position and size) and two non-Euclidean features (image statistics and spatial frequency content of an image). Overall, identity representation increased and nonidentity feature representation decreased along the ventral visual pathway, with identity outweighed the non-Euclidean features, but not the Euclidean ones, in higher levels of visual processing. A similar analysis was performed in 14 convolutional neural networks (CNNs) pretrained to perform object categorization with varying architecture, depth, and with/without recurrent processing. While the relative coding strength of object identity and nonidentity features in lower CNN layers matched well with that in early human visual areas, the match between higher CNN layers and higher human visual regions were limited. Similar results were obtained regardless of whether a CNN was trained with real-world or stylized object images that emphasized shape representation. Together, by measuring the relative coding strength of object identity and nonidentity features, our approach provided a new tool to characterize feature coding in the human brain and the correspondence between the brain and CNNs.

 bioRxiv Subject Collection: Neuroscience

 Read More

Leave a Reply

%d bloggers like this: