"Can Deep Neural Networks Model the Robustness of Human Object Recognition?"
Frank Tong, Ph.D.
The goal of Frank Tong's research is to investigate, characterize, and model the neural mechanisms that mediate human visual perception and cognition. What allows us to detect the presence of a clump of weeds in a lawn, to recognize an animal hiding behind a bush, or to remember the precise hue and texture of an ocean surface during sunset? A core assumption in his work is that early visual representations have a powerful but under appreciated role in higher cognitive operations, and that higher-level mechanisms of attention and working memory serve to modulate processing at early visual sites to select and maintain task-relevant visual information. Characterizing and modeling the interplay between early visual representations and higher order representations represents a long-term goal of this work. The work relies on behavioral and psychophysical methods, high-resolution fMRI, and advanced computational approaches for both data analysis and modeling. The lab has developed novel methods for decoding feature-selective responses from patterns of fMRI activity in the human visual cortex (Kamitani & Tong, Nature Neuroscience, 2005; Current Biology, 2006; Tong & Pratte, Annual Review of Psychology, 2012), and shown how these approaches can be used to characterize the neural bases of visual working memory (Harrison & Tong, Nature, 2009; Pratte et al., 2014) and object-based attentional selection (Pratte et al., J Neurophysiology, 2013; Cohen & Tong, Cerebral Cortex, 2015). In ongoing work, the Tong lab is developing, training, and testing deep convolutional neural networks as potential models for understanding the neural bases of human visual processing.