LEARNING AND EXPLICITATION OF GRADUAL RULES USING ARTIFICIAL NEURAL NETWORKS

摘要

This work belongs to the field of neural architectures and hybrid systems for Artificial Intelligence (AI). It concerns the study of "gradual" rules, which makes it possible to represent correlations and modulation relations between variables. We propose a set of characteristics to identify these gradual rules, and a classification of these rules into "direct" rules and "modulation" rules. In neurobiology, pre-synaptic neuronal connections lead to gradual processing and modulation of cognitive information. While taking as a starting point such neurobiological data, we propose in the field of connectionism the use of "Sigma-Pi" connections to allow gradual processing in AI systems, In order to represent as well as possible the modulation processes between the inputs of a network, we have created a new type of connection, "Asymmetric Sigma-Pi" (ASP) units. They facilitate the extraction or explicitation of gradual rules from a neural network. The method suggested is based on the analysis of the derivative of the output of the network compared to the inputs connected to an ASP unit. These models have been implemented within a pre-existing hybrid neuro-symbolic system, the INSS system, based on connectionist nets of the "Cascade Correlation". type. The new hybrid system thus obtained, INSS-Gradual, allows the learning of bases of examples containing gradual modulation relations.