Introduction aux réseaux conceptuels appliqués à l’apprentissage automatique des machines

Abstract

This thesis introduces a machine learning model, named conceptual network, to encode the sensory experiences of a computer system ; with the different variants of its units (premises, concepts, structural concepts, homogeneous concepts, heterogeneous concepts), as well as the mechanisms necessary for the generation and exploitation of the network units : conceptualization, genesis, activation and evocation mechanisms. Formed from interconnected units, called concepts, the conceptual network is automatically constructed from raw sensory data according to a complexity (number of units and relations between units) that tends to follow that of sensory phenomena. In doing so, the structure of the network does not model the input/output relations (contrary to classical approaches : neural networks, SVM, decision trees, Bayesian networks), but rather the observed phenomena as such. The conceptual network is thus intended to encode the phenomenological universe of the system from which it is possible to develop several strategies for pattern recognition, regression or possibly the control of robotic systems. For example, for the prediction objective, a concept-label for each class can be added to the sensory inputs of the network. In a new experiment, where the concepts that encode the structure of the observations are activated more in particular, it is possible to predict the class by comparing the energies of the different concept-labels. Based on this principle, we present some preliminary examples of conceptual networks used for digit recognition (written and spoken) and for speaker recognition. The advantage of the proposed paradigm, compared to the traditional input/output paradigm, is firstly that the model thus realized can be used for any task without requiring significant adaptation of the structure and parameters of the network, not being optimized for specific tasks. Second, the automatic construction of the network allows to reach a structural complexity that is limited to that of the observed phenomena : with, potentially, millions of units and billions of connections. In contrast, encoding phenomenological experiences of unknown (and potentially dynamic) complexity would be more difficult with models whose structure is fixed ; if that were the objective of these models. Third, conceptual networks have the particularity of learning quickly with few examples. For large corpora, managing the learning rate (and the inherent complexity of the structure) is a challenge at the moment. Limitations and future perspectives are discussed.

Date	2 Nov 2023
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Patrick Cardinal (Supervisor) & Pierrich Plusquellec (Co-supervisor)