Integrating Computational And Neural Findings In Visual Object Perception PDF Download

There is a promise at the heart of neuroscientific research: understanding the brain will provide insights that improve health and well being. To fulfill this promise, our theories of neural function must be able to operate `at scale'--i.e., generalize to naturalistic environments. However, even in well-controlled experimental settings, neuroscientific theories often fail to replicate, much less generalize. Recent advances in computer science have created computational frameworks which can operate at scale and necessarily generalize across experiments. But leveraging these methods to understand neural function beyond canonical visual cortex has had limited success. In this dissertation I demonstrate how these methods can be used to formalize and evaluate theories of neural function downstream of canonical sensory cortex. First, I review the historical challenges in understanding perirhinal cortex (PRC), including a longstanding debate over PRC involvement in perception (chapter 1). To formalize and evaluate PRC involvement in visual object perception, I situate lesion, electrophysiological, and behavioural data within a deep learning computational framework. This work resolves decades of apparent inconsistencies in the experimental literature by integrating results from human (chapter 2) and monkey (chapter 3) experimental data within a `stimulus-computable' modeling framework. I extend this work to better isolate and characterize PRC contributions to visual object perception (chapter 4). Finally, I synthesize neuroscientific findings from multiple species to provide an account of PRC involvement in visual perception (chapter 5). Taken together, these data suggest that PRC supports visual object perception by integrating over the sequential outputs of canonical visual cortices. I conclude by describing how, to further understand how PRC supports these behaviors, this neuroscientific data can be situated within a biologically plausible optimization framework: we are well-positioned to understand perirhinal function using optimization approaches from computer science, constrained to reflect what we know about the underlying biology. This approach promises not only to formalize the relationship between experimental variables and PRC function, but to instantiate our theories of neural function in a way that is designed to operate at scale.