Learning from Imperfect Supervisor Using Neural Network Ensemble (特集:コンピュータグラフィクスの新展開)

概要

論文の詳細を見る
In training a conventional multilayer perceptron (MLP), the existence of perfect learning data or a perfect supervisor usually has to be assumed. But in on-line training processes in which a human expert has no direct access to the training data, it is probable that erroneous data due to imperfection of the learning object (supervisor) or observation noise will be included in the training data. This kind of learning data will cause a non-optimal learning process that produces an unreliable neural network. In this paper we propose a model of a neural network ensemble consisting of a number of MLPs in which each MLP can be trained to obtain a unique type of expertise. On the basis of each MLP's expertise, the ensemble will be able to distinguish correct learning data from erroneous data and automatically assign one of its members to be trained only on correct data, allowing an optimal learning process with an imperfect learning supervisor. The proposed ensemble model, eliminates the requirement to design a perfect learning supervisor to train neural networks, and can thus help to make neural networks more widely applicable in real-world problems.
一般社団法人情報処理学会の論文
2001-05-15