Parametric activation functions modelling fuzzy connectives for better explainability of neural models
Luca Sára Pusztaházi, Gábor Csiszár, Michael S. Gashler, Orsolya Csiszár · 2022
In this work, a neuro-fuzzy hybrid deep learning model is presented for finding human-readable relationships between input features with the help of nilpotent fuzzy logic and multi-criteria decision making (MCDM). In the neural network a parameterized, differentiable activation function is used, where the parameter is determined by gradient descent. The goal is to find the optimal regularization value by applying the deep learning model to classification problems from the UCI Machine Learning Repository.