Technical Program

Paper Detail

Paper: GS-4.2
Session: Contributed Talks 7-8
Location: H0104
Session Time: Sunday, September 15, 11:50 - 12:30
Presentation Time:Sunday, September 15, 12:10 - 12:30
Presentation: Oral
Publication: 2019 Conference on Cognitive Computational Neuroscience, 13-16 September 2019, Berlin, Germany
Paper Title: Why Are Face and Object Processing Segregated in the Human Brain? Testing Computational Hypotheses with Deep Convolutional Neural Networks
Manuscript:  Click here to view manuscript
View Video: Video
License: Creative Commons License
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Authors: Katharina Dobs, Massachusetts Institute of Technology, United States; Alexander Kell, Columbia University, United States; Ian Palmer, Michael Cohen, Nancy Kanwisher, Massachusetts Institute of Technology, United States
Abstract: Why does the human brain contain cortical regions specialized for the perception of some stimulus categories (e.g., faces), but not others (e.g., cars)? And why might functional specialization be a good design strategy for brains in the first place? Here, we used deep convolutional neural networks (CNNs) to test whether models optimized to recognize faces and objects require functional segregation for each task. First, we trained two separate CNNs with the same architecture to categorize either faces or objects. Unsurprisingly, the face-trained CNN performed worse on object categorization than the object-trained CNN and vice versa, demonstrating that the features optimized for each task differ from one another. Second, following the method of Kell et al (2018), we trained a family of dual-task CNNs on both tasks, asking how many layers can be shared before performance declines. Somewhat surprisingly, even the dual-task CNN that shared all layers performed nearly as well as the separate networks. This result is consistent with two hypotheses: 1) face and object recognition may be performed well by using a shared pool of common features or 2) the shared network has learned “hidden” functional specialization. In ongoing work, we are seeking to disambiguate these two hypotheses.