Presentation | 2014-03-14 Experimental Study on Effect of Pre-training in Deep Learning through Visualization of Unit Outputs Tsubasa OCHIAI, Hideyuki WATANABE, Shigeru KATAGIRI, Miho OHSAKI, Shigeki MATSUDA, Chiori HORI, |
---|---|
PDF Download Page | PDF download Page Link |
Abstract(in Japanese) | (See Japanese page) |
Abstract(in English) | To clarify the capability of recent powerful classifier concept, Deep Neural Networks (DNN), we experimentally investigate effects of the pre-training used to initialize DNN. A deep neural network is first pre-trained using Restricted Boltzmann Machine (RBM), then it is run as an embodiment of Deep Belief Networks, which basically possess associative memory function, and a Deep Autoencoder, which is expected to realize feature representation for an input pattern over the inner layers of network. Analyses are conducted through the visualization of network unit outputs. Based on the experiments, we reveal that the RBM-based pre-training successfully makes networks memorize some information of training patterns and also represent pattern features inside the networks. |
Keyword(in Japanese) | (See Japanese page) |
Keyword(in English) | Deep Learning / Pre-training / Visualization of unit outputs / Deep Neural Networks |
Paper # | PRMU2013-210 |
Date of Issue |
Conference Information | |
Committee | PRMU |
---|---|
Conference Date | 2014/3/6(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 | Pattern Recognition and Media Understanding (PRMU) |
---|---|
Language | JPN |
Title (in Japanese) | (See Japanese page) |
Sub Title (in Japanese) | (See Japanese page) |
Title (in English) | Experimental Study on Effect of Pre-training in Deep Learning through Visualization of Unit Outputs |
Sub Title (in English) | |
Keyword(1) | Deep Learning |
Keyword(2) | Pre-training |
Keyword(3) | Visualization of unit outputs |
Keyword(4) | Deep Neural Networks |
1st Author's Name | Tsubasa OCHIAI |
1st Author's Affiliation | Doshisha University() |
2nd Author's Name | Hideyuki WATANABE |
2nd Author's Affiliation | NICT:Doshisha University |
3rd Author's Name | Shigeru KATAGIRI |
3rd Author's Affiliation | Doshisha University |
4th Author's Name | Miho OHSAKI |
4th Author's Affiliation | Doshisha University |
5th Author's Name | Shigeki MATSUDA |
5th Author's Affiliation | NICT |
6th Author's Name | Chiori HORI |
6th Author's Affiliation | NICT |
Date | 2014-03-14 |
Paper # | PRMU2013-210 |
Volume (vol) | vol.113 |
Number (no) | 493 |
Page | pp.pp.- |
#Pages | 6 |
Date of Issue |