Technical Program

Paper Detail

Paper: PS-1B.6
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: A deep generative model explaining tuning properties of monkey face processing patches
Manuscript:  Click here to view manuscript
License: Creative Commons License
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Authors: Haruo Hosoya, ATR International, Japan
Abstract: Recent monkey studies have revealed a face processing network in the IT cortex that consists of multiple face-selective patches and forms a putative functional hierarchy. Although a number of computational models accounting for this have been proposed, they have been mostly feedforward, ignoring the reciprocal nature of the visual system. Here, we present a two-layer deep generative model based on variational autoencoder (VAE), which provides a Bayesian probabilistic framework with explicit feedforward and feedback processing. While the lower layer of our model uses a standard VAE, the upper layer uses our recently developed algorithm called group-based VAE, which is capable of learning invariant representations from inputs with grouping information. After training with multi-view face images, the upper layer encoded view-invariant facial identities while the lower layer showed facial feature tuning, both in a way quantitatively similar to the observations in patches AM and ML, respectively, as shown in Freiwald and Tsao (2010) and Freiwald et al. (2009). Taken together, we have found a novel deep generative model that might have some computational relevance with the monkey face processing system.