Associative Memory in a Quantum-Optical Neural Network

We present a quantum-optical neural network made of ultracold atoms and photons that functions as an associative memory. The neurons are realized by an array of 16 Bose-Einstein condensates and are connected by superpositions of degenerate light fields in a multimode optical cavity. As an associative memory, the neural network is able to recall stored patterns even when presented with missing or distorted information. This functionality was first proposed in the Hopfield neural network, the subject of the 2024 Nobel prize. Hopfield networks function well for recalling a small number of memories, but when overloaded with too many they break down and become something quite different: a spin glass. These are highly frustrated networks characterized by disorder. While remarkable in their own right, as recognized by the 2021 Nobel prize, spin glasses were thought to be devoid of associative memory with extensive recall. Here, we provide a counterexample: the quantum-optical neural network operates in a spin glass phase and leverages the native driven-dissipative, open quantum dynamics to realize an associative memory with greater memory capacity than Hopfield networks. While standard 16-neuron Hopfield networks reliably only store 2-3 memories, we have discovered quantum-optical neural networks of the same size that store up to 8 memories with greater than 80% recall probability. Harnessing quantum physics for neural networks could enable a new class of neuromorphic quantum devices that go beyond the capabilities of gate-based quantum computers or classical neural networks alone. Our results take a first step in unlocking this potential and present a clear path to a fully quantum limit.