Technical Program

Paper Detail

Paper: PS-1B.32
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: Modeling human visual responses with a U-shaped deep neural network for motion flow-field estimation
Manuscript:  Click here to view manuscript
Authors: Atsushi Wada, Satoshi Nishida, Hiroshi Ando, Shinji Nishimoto, National Institute of Information and Communications Technology, Japan
Abstract: Deep Neural Networks (DNNs) for visual classification have recently been shown to exhibit representations homologous to those observed in the human visual system. For further exploring the relationship between artificial and natural neural networks, we focus on a U-shaped contraction-expansion architecture for DNNs, which recursively refine the network output for solving pixel-wise 2D-prediction tasks. By using FlowNet (Dosovitskiy et al., 2015), a U-shaped DNN for motion flow-field estimation, we show that the DNN-features extracted from FlowNet accurately predict human visual responses to natural movie stimuli. We further present that the features from the expansion compared to contraction layers yield higher encoding performance for the mid-level regions in the dorsal visual pathway. Our results may support the notion of the information integration between the early processing stages preserving fine spatial information and the downstream processing stages providing global contextual cues for motion estimation in the human visual system.