Optimizing Composite Neural Networks for Very Hard Classification Problems

概要

論文の詳細を見る
In this paper, we discuss how to determine the convergence of training and the generalization capability of single neural networks. and how to improve recongition rates by using composite neural networks for very hard classification problems. First, the input region of each class is approximated by hyperboxes, which are recursively defined by resolving overlaps between classes. The level of recursion between two classes indicates the difficulty of separating them. Using the level of recursion for training data, we can determine which classes are difficult to separate and then using that for as test data, we can check the generalization ability. Next, using the level of recursion, we group into the same superclass the classes that are difficult to separate. The first-level neural network in the composite neural network classifies the input data into a superclass, and the second-level neural networks separate the input data into classes. We apply our method to blood cell classification, which is known to be a very hard problem. The results show that our method requires less time than that required for training a conventional network, and has the additional advantage of improved recognition rates.
一般社団法人情報処理学会の論文
1994-09-15