Associative memory as a Bayesian building block

Learning in neural systems relies on forming associative memories through changes in synaptic connections, enabling information retrieval via learned associations. This thesis examines the Palm network as a biologically inspired associative memory model that approximates Bayesian inference in a distributed manner. We analyze its computational properties, including scalability, stability, fault tolerance, and parallel processing, and interpret its operation as a Voronoi‑based classifier. Extensions to weighted regions, reinforcement learning, and spiking neuron implementations are explored. Building on this analysis, we propose a Bayesian Memory framework that supports hierarchical organization, bidirectional information flow, and efficient data reduction, providing a theoretical foundation for scalable associative memory networks capable of supporting cognitive‑level processing.