Apprentissage continu bio-inspiré dans les réseaux à mémoire associative

Continuous learning, the ability to acquire new memories and skills while preserving previously learned information, is a fundamental capability of the brain. However, replicating this capability in artificial neural networks remains problematic, as they typically suffer from catastrophic forgetting (CF), the abrupt and uncontrolled loss of previously stored information when learning new data.This thesis investigates continuous learning and catastrophic forgetting dynamics in associative memory networks (AMNs), computational models proposed as analogues of hippocampal circuits due to their recurrent architecture and attractor dynamics. Specifically, we study continuous learning in the challenging scenario of sequential storage of highly correlated memory patterns. These patterns rely on overlapping neural representations, where the same groups of neurons participate in encoding multiple memories. Such overlapping representation is considered a fundamental organizational principle of the brain, enabling generalization and the extraction of regularities across experiences. However, in AMNs, the sequential storage of overlapping memory representations typically induces interference between stored memories, ultimately resulting in catastrophic forgetting.We introduce a novel mechanism that mitigates catastrophic forgetting in AMNs by enabling autonomous rehearsal of stored patterns. By incorporating biologically inspired plastic self-inhibition into continuous Hopfield networks, we demonstrate that the network can spontaneously explore its attractor landscape, sequentially retrieving all previously stored memories without external intervention. Using a gradient descent algorithm, the rehearsal mechanism allows the network to construct synaptic configurations that simultaneously support both old and new memories. By limiting interference between stored memories, this mechanism prevents catastrophic forgetting during the sequential incorporation of new correlated patterns.Our approach satisfies key biological constraints: all learning rules are local, relying only on information available at individual synapses, and the system operates autonomously without requiring an external catalog of previously stored patterns. The resulting framework scales effectively with network size and successfully handles patterns with varying degrees of correlation.This work aims to contribute to both neuroscience and artificial intelligence by proposing a biologically plausible mechanism for continuous learning that could inform our understanding of memory consolidation in the brain and guide the development of more robust artificial learning systems, particularly for neuromorphic implementations where local computation is essential.