Selection and Re-learning of learning data based on quantization of weights in multi-layered neural networks

摘要

In this report, methods for selection, re-learning of learning data, and structurization of multilayered neural networks(NNs) based on the quantization procedure of weights in NNs are newly propoed. Here an idea concerning with the balance between the size of the learning data and the size of a NN is introduced. In the proposed learning algorithm, by using the following procedures; (i)the reduction of the number of the quantization bits and (ii)the limitation of the learning data by the quantization processing, the structurization of NNs can be realized. Finally, we apply the proposed learning algorithm for classfication problems and show its structurization and task-decomposition abilities concretely.