Search Papers | Poster Sessions | All Posters
Poster B67 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
A single computational objective may not be sufficient for human-like face discrimination
Ammar Marvi1 (), Chengxu Zhuang1, Katharina Dobs2, Nancy Kanwisher1; 1Massachusetts Institute of Technology, 2Justus-Liebig University Giessen, Hesse, Germany
Category-selective cortical regions are key to the diverse capabilities supported by the ventral visual pathway (VVP). How might these regions arise in development? One hypothesis suggests that category selectivity emerges from domain-general learning mechanisms. Supporting this idea, artificial neural networks (ANN) trained with self-supervised learning objectives on natural images recapitulate many features of the functional organization of the VVP and show human-like visual classification abilities. However, an adequate model of VVP development should account for all the behavioral abilities it supports, including face recognition. We therefore trained and tested a set of self-supervised networks on datasets composed of object or face images. When testing models via a naturalistic, zero-shot task we find that object-trained models achieve human-level accuracy in the object recognition task, yet face-trained models fail catastrophically at face discrimination. However, when provided labels after training and assessed via linear readout, all models - including those trained on faces - yield high discrimination accuracy, approaching performance of their supervised counterparts. Thus human-like face recognition may not develop from domain-general learning mechanisms. Instead, a single computational objective may only suffice if given a prior on the number, grain, or identity of output categories.
Keywords: face discrimination self-supervised learning objectives neural organization deep neural networks