Neural network explanation using inversion

Abstract

An important drawback of many artificial neural networks (ANN) is their lack of explanation capability [Andrews, R., Diederich, J., & Tickle, A. B. (1996). A survey and critique of techniques for extracting rules from trained artificial neural networks. Knowledge-Based Systems, 8, 373-389]. This paper starts with a survey of algorithms which attempt to explain the ANN output. We then present HYPINV, a new explanation algorithm which relies on network inversion; i.e. calculating the ANN input which produces a desired output. HYPINV is a pedagogical algorithm, that extracts rules, in the form of hyperplanes. It is able to generate rules with arbitrarily desired fidelity, maintaining a fidelity-complexity tradeoff. To our knowledge, HYPINV is the only pedagogical rule extraction method, which extracts hyperplane rules from continuous or binary attribute neural networks. Different network inversion techniques, involving gradient descent as well as an evolutionary algorithm, are presented. An information theoretic treatment of rule extraction is presented. HYPINV is applied to example synthetic problems, to a real aerospace problem, and compared with similar algorithms using benchmark problems.