Technical Program

Paper Detail

Paper: PS-1B.74
Session: Poster Session 1B
Location: H Fl├Ąche 1.OG
Session Time: Saturday, September 14, 16:30 - 19:30
Presentation Time:Saturday, September 14, 16:30 - 19:30
Presentation: Poster
Publication: 2019 Conference on Cognitive Computational Neuroscience, 13-16 September 2019, Berlin, Germany
Paper Title: Time-Resolved Correspondences Between a Deep Feed-Forward Neural Network and Human Object Processing: EEG Measurements
Manuscript:  Click here to view manuscript
License: Creative Commons License
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
DOI: https://doi.org/10.32470/CCN.2019.1082-0
Authors: Nathan Kong, Blair Kaneshiro, Anthony Norcia, Stanford University, United States
Abstract: The ventral visual system is known to exhibit hierarchical structure, where early and higher visual areas respond to simple and relatively complex features respectively. A clear, quantitative explanation for the image computations performed in each visual area is, however, lacking. Feed-forward hierarchical convolutional neural networks have been a step forward in attempting to model these computations. Here we model the temporal evolution of EEG responses recorded during passive viewing of multiple object categories using layers of a convolutional neural network trained to perform image categorization. We found a modest hierarchical correspondence between the depth of the layer in the neural network and the neural response time at which model and neural representations are maximally correlated. However, we show that shallow layer and deep layer representations start to correlate with neural representations at similar time bins. A reliability analysis indicated that the modest correspondences are far from the limit imposed by variability of the data, but are largely due to the inadequacies of the model. These results provide suggestive evidence that early visual areas perform more than just simple feature detection and that strictly feed-forward convolutional neural network models are insufficient to model human object processing dynamics.