Progressive feature extraction by extended greedy information acquisition
Ryotaro Kamimura, Haruhiko Takeuchi, Osamu Uchida · 2004
We propose a new network growing method to detect salient features in input patterns. The new method is based upon a previous network growing model (R. Kamimura and T. Kamimura, 2002) and introduced to overcome some problems in the previous model. We have so far tried to build a model that can learn input patterns as efficiently as possible. To realize this efficiency, we impose upon networks a constraint that only connections into new competitive units must be updated to absorb as much information as possible from outside. However, one of the problems is that the previous improper feature extraction prevents networks from extracting appropriate features in the later learning stages. To overcome this problem, we relax the condition of the previous model, and we permit networks to update all connections for gradual feature extraction at the expense of computational efficiency. We applied the new method to a simple problem that the previous model cannot solve, and information education data analysis. In both problems, we found that the new method can appropriately extract features from input patterns.