Learning with noise in engineering domains

Abstract

This paper presents a learning approach which identifies and eliminates noisy examples (outliers) to improve the quality of training on engineering data and the effectiveness of the learned concept descriptions. In this approach, one (1) acquires initial concept descriptions from preclassified attributional training data, (2) optimizes concept descriptions to improve their descriptiveness, (3) applies optimized concept descriptions to filtrate/improve initial training data, and (4) repeats the learning process from improved training data. The implemented algorithm extends the widely used open loop learning approach (divided into concept acquisition phase and concept optimization phase) into a closed loop learning approach. In the closed loop learning approach, learned and optimized concept descriptions are fed back and used to filter training data for the next learning iteration. Thus, the learning program is run at least two times; the first time to acquire concept descriptions for the optimization step, and the second time to acquire the final descriptions. In this approach, noise is detected on the concept description level rather than on the raw data level — where the evaluation of raw data can be impossible since the training data may be composed of numeric, symbolic, relational and structural attributes. This method is successfully applied to different engineering problems, and its effectiveness is illustrated for three qualitatively different problems in computer vision.