A new method for explaining neural network reasoning

This paper presents a new method for explaining the reasoning results of a trained neural network. The method considers the most significant attribute first under the guidance of a relative strength of effect analysis and eliminates irrelevant points. Following the adaptive search in the dynamic state space, a set of relevant points are extracted and form the basis of the explanation of the neural network reasoning. Combining a relative strength of effect analysis with the relevant points, a case based explanation approach is put forward. As an illustration, an experiment with a small data set on the relationship between weather conditions and play decisions is presented to demonstrate the utility of the proposed approach.