Neural Network Hardware Eliminating Negligible Connections
-
- HOSHI Noritaka
- School of Engineering, Nagoya University
-
- ICHIKAWA Shuichi
- School of Engineering, Nagoya University
-
- SHIMADA Toshio
- School of Engineering, Nagoya University
Bibliographic Information
- Other Title
-
- 不要な結合重みを削除するニューラルネットワーク専用ハードウェア
Search this article
Abstract
This Paper proposes a new hardware architecture for neural network, in particular, sparsely connected. neural network. In study mode, this hardware checks weights which are small enough (nearly zero) , and eliminates negligible connections. Study is gradually accelarated as connection eliminated. It is very flexible and applicable to very large and irregularly structured neural network. Simulations show that eliminating negligible connections is effective in shortening of runtime.
Journal
-
- IEICE technical report. Computer systems
-
IEICE technical report. Computer systems 95 (502), 57-64, 1996-01-31
The Institute of Electronics, Information and Communication Engineers
