Fuzzy Systems with Sigmoid-Based Membership Functions as Interpretable Neural Networks

Abstract

In this paper new interpretable neural network architectures are proposed. A Neural Network with sigmoid activation function is converted into a sigmoid-based approximation operator, which, at its turn, can be approximated by a fuzzy system of Takagi-Sugeno type. Altogether this process shows that a neural network with sigmoid activation functions can be approximated by a Takagi-Sugeno fuzzy system. As the TS fuzzy system provides interpretability, while Neural Networks provide approximation capability, we obtain a novel interpretable Neural Network Architecture. Interpretability of the TS fuzzy system can provide insights into data in a new way, enhancing decision making.