| Presentation | 2007-03-16 Learning nonlinear forward optics in generative models Satohiro TAJIMA, Masataka WATANABE, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | Visual processing is an inverse problem, say, the system needs to unravel the three-imensional representation of the world relying on two-dimensional retinal input. Although the generative model is a promising method, acquisition of complex forward optics remains a problem. Here we propose a solution to this problem using a multilayered neural network with backpropagation learning. Unlike conventional applications of multilayered network to vision, we assume that inputs are high-level object representations, whereas the desired output is low level retinal representation. We chose "occlusion" as an example nonlinear forward optics and simulated the learning process in accordance to the findings that infants learn to recognize occlusion at a later stage of visual development. We incorporated this network into a generative model and simulated the estimation process of object depth when given partially occluded multiple objects. Analysis of hidden units revealed that there exists various type of neural preference depending on the luminance of occluded and non-occluded surface of visual objects. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | Generative Model / Top-Down / Visual Processing / Occlusion / Back-Propagation |
| Paper # | NC2006-188 |
| Date of Issue |
| Conference Information | |
| Committee | NC |
|---|---|
| Conference Date | 2007/3/9(1days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | |
| Chair | |
| Vice Chair | |
| Secretary | |
| Assistant | |
| Paper Information | |
| Registration To | Neurocomputing (NC) |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | Learning nonlinear forward optics in generative models |
| Sub Title (in English) | |
| Keyword(1) | Generative Model |
| Keyword(2) | Top-Down |
| Keyword(3) | Visual Processing |
| Keyword(4) | Occlusion |
| Keyword(5) | Back-Propagation |
| 1st Author's Name | Satohiro TAJIMA |
| 1st Author's Affiliation | Faculty of Engineering, University of Tokyo() |
| 2nd Author's Name | Masataka WATANABE |
| 2nd Author's Affiliation | Faculty of Engineering, University of Tokyo |
| Date | 2007-03-16 |
| Paper # | NC2006-188 |
| Volume (vol) | vol.106 |
| Number (no) | 590 |
| Page | pp.pp.- |
| #Pages | 6 |
| Date of Issue | |