Résumé
In this paper, we propose a new learning method for max-min neural networks with more than three layers. These networks, valued for their intrinsic interpretability due to their direct link with fuzzy logic if-then rules, face computational challenges stemming from the non-linearity of max-min operations. Unlike traditional gradient descent approaches, which are ineffective in this context (owing to zero derivatives for most synaptic weights), the proposed learning method relies on approximate and stable solutions to the max-min equations. This method uses a three-step backpropagation for each neuron layer: an update of the synaptic weights by solving a max-min equation followed by an update using an approximate gradient and then a backpropagation of the synaptic values by solving another max-min equation. Experiments on synthetic and real-world data show promising performance. This method paves the way for interpretable neural networks that can be used in complex scenarios.