Investigating the Behavioral, Neural, and Computational Mechanisms of Object Recognition and Memorability Under Temporal Constraints

Abstract

The primate inferior temporal (IT) cortex transforms retinal inputs into object-centered representations supporting visual recognition. Recent work suggests IT also predicts image memorability, yet IT activity disappears within hundreds of milliseconds after stimulus offset, raising questions about how information is maintained during delays.

This thesis investigates IT's contributions to recognition and memory. In Aim 1, we combined human behavior, macaque neurophysiology, and neural network modeling to test whether optimizing networks for recognition and memorability improves correspondence with IT. Jointly trained models explained more neural variance and better predicted both behavioral measures. In Aim 2, we examined these representations in a delayed match-to-sample task. Behavioral accuracy declined with increasing delay despite IT activity vanishing, suggesting that maintaining object-selective representations requires transformation of transient IT codes into persistent states in downstream circuits.

These findings show IT encodes representations optimized for multiple functions, while maintenance across time depends on recurrent processing beyond IT.