Extracting fuzzy symbolic representation from artificial neuralnetworks

摘要

The paper presents FUZZYTREPAN, a pedagogical approach to theproblem of extracting comprehensible symbolic knowledge from trainedartificial neural networks. This approach extends the previouslyproposed TREPAN method in two ways: it uses fuzzy representation in itsknowledge extraction process (by means of fuzzy decision trees), and ituses additional heuristics in its process of generating artificial data.The paper describes the proposed approach in detail, and it presents itsempirical evaluation on popular machine learning benchmarks