New Learning Rules for Hierarchical Neural Networks for Visual Pattern Recognition

2017 Fiscal Year Final Research Report

• Project/Area Number: 25330300
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Research Field: Soft computing
• Research Institution: Fuzzy Logic Systems Institute
• Principal Investigator: FUKUSHIMA KUNIHIKO 一般財団法人ファジィシステム研究所, 研究部, 特別研究員 (90218909)
• Project Period (FY): 2013-04-01 – 2018-03-31
• Keywords: 視覚パターン認識 / 多層神経回路 / deep CNN / ネオコグニトロン / 学習手法 / 内挿ベクトル法 / Add-if-Silent / margined WTA

Outline of Final Research Achievements

We developed new learning rules for the neocognitron, which is a deep convolutional neural network for visual pattern recognition.
For training intermediate layers, the learning rule named AiS (Add-if-Silent) is used. Under the AiS, a new cell is generated and added to the network if all postsynaptic cells are silent in spite of non-silent presynaptic cells. Once a cell is generated, its input connections do not change any more. Thus the training process is very simple and does not require time-consuming repetitive calculation.
For training the deepest layer, we proposed a supervised learning rule called mWTA (margined Winner-Take-All). Every time when a training pattern is presented during the learning, if the result of classification by the WTA is an error, a new cell is generated. Here we put a certain amount of margin to the WTA. The mWTA produces a compact set of cells, with which a high recognition rate can be obtained with a small computational cost.

Free Research Field

人工神経回路