Technical Program

Paper Detail

Paper: PS-1B.24
Session: Poster Session 1B
Location: Symphony/Overture
Session Time: Thursday, September 6, 18:45 - 20:45
Presentation Time:Thursday, September 6, 18:45 - 20:45
Presentation: Poster
Paper Title: Using features from deep neural networks to model human categorization of natural images
Manuscript:  Click here to view manuscript
Authors: Ruairidh Battleday, Joshua Peterson, Thomas Griffiths, University of California, Berkeley, United States
Abstract: An open question is whether recent advances in machine learning can be used to study human intelligence. Recent work from neuroscience has shown that stimulus representations from convolutional neural networks (CNNs) are our best currently available predictors of several visual areas in the brain, and behavioral studies have found they best explain human similarity judgments of natural object images. A prime candidate to continue this avenue of investigation is categorization, a fundamental cognitive function. Although a range of high-precision formal models of categorization exist, they are limited to simple, artificial stimuli because representing more realistic stimuli is difficult. We show that representations derived from CNNs can be used to extend these models to natural images, and that a group of cognitive models based on these representations capture human behavior over a novel crowdsourced database of >500,000 human classification decisions. Successful models include both exemplar and prototype models, contrasting with the dominance of exemplar models in previous work. We also find that performance is improved by using representations from networks that score more highly at ground-truth prediction.