A Noise-Tolerant Approach to Symbolic Learning from Sensory Data

Abstract

This article presents a novel approach to noise-tolerant symbolic learning. The approach is applicable to many existing learning programs, and it has been implemented for the AQ-14 (rule learning) and C4.5 (decision tree learning) programs. In this approach, the system (1) acquires initial concept descriptions from pre-classified attributional training data, (2) optimizes concept descriptions to improve their descriptiveness, (3) applies optimized concept descriptions to filtrate initial training data, and (4) repeats the learning process from filtered data. This method outperforms the traditional open-loop, single step learning procedure when applied to the texture recognition problem of 12 classes and to the image annotation problem of natural scenes.