Beyond hebbian learning: optimizing memory retrieval and network structure in bio-inspired neural models

This thesis investigates how memory storage and retrieval can be optimized in bio-inspired neural networks by going beyond classical Hebbian learning and fully symmetric architectures. In the first part, we introduce the Daydreaming algorithm, a local and unsupervised learning rule that combines learning and unlearning to stabilize stored memories while suppressing spurious attractors. This approach overcomes the classical capacity limit of the Hopfield model, achieving optimal performance for uncorrelated patterns and efficiently exploiting correlations in structured data, including real-world datasets. In the second part, we explore neural networks with asymmetric and diluted connectivity, closer to biological systems. By systematically studying their dynamical behavior, we identify connectivity regimes that maximize attractor diversity, robustness, and computational efficiency. Together, these results highlight the crucial role of both synaptic plasticity and network structure in associative memory systems.