Associative Memory Deep Dive

A four-part interactive deep dive into associative memory for superconducting optoelectronic networks. Part one explores the storage capacity of a Hopfield network and its catastrophic breakdown past a load of about 0.138 patterns per neuron. Part two visualizes the energy landscape and basins of attraction in a two-pattern overlap plane, with a state rolling downhill to the nearest stored memory. Part three maps the Hopfield network onto superconducting hardware, where symmetric weights are mutual inductances between flux loops and spurious memories are spin-glass frustration. Part four shows the modern dense Hopfield network, whose sharper energy wells give exponential capacity and whose retrieval rule is exactly the attention mechanism of transformers.