2026-09-22
A 16-spin atom-cavity glass recalls memories via nonequilibrium descent. Mean capacity is 11.9; softer traps hit 25, about 7× a finite-size Hebbian Hopfield net.
The Hopfield net stores memories as energy minima of an all-to-all Ising system. Hebbian synapses write the patterns; equilibrium Metropolis-Hastings (MH) dynamics does pattern completion. Amit et al. put a thermodynamic cap at P0 ≈ 0.14n. Past that point the landscape becomes a spin glass: exponentially many spurious minima, nested valleys, high barriers. A single spin-flip error can dump the state into a neighboring basin, so those minima cannot serve as memories. Glasses were long treated as incompatible with associative recall.
That cap is a real constraint for physical hardware. Dense associative memories raise capacity by engineering interactions of order higher than two, a construction later tied to Transformers, but they do not use the glass. In 2021 the Stanford Lev and Ganguli group argued that driven-dissipative cavity QED would implement steepest-descent spin flips, with a rate proportional to the energy dropped, enlarge the basins of spurious minima, and turn the glass into a memory. This Science paper builds that network in hardware at up to 16 spins, and finds a second boost nobody had put in the theory: atoms move, Jij changes, and the landscape elastically deepens the basin the system is already in.
The hardware is an array of rubidium Bose-Einstein condensates in a multimode optical cavity. Each tweezer holds about 4.2×10⁴ atoms. Two checkerboard density waves play the role of spin up and down; cavity photons supply all-to-all, sign-changing synapses Jij. A "4/7" cavity geometry (mirror spacing over radius of curvature ≈ 1.22) reduces the physics to a frustrated transverse-field Ising model with local bias fields. A digital micromirror device splits a longitudinal pump into 16 beams, each imposing a stimulus field fi, i.e. a possibly corrupted input pattern.
Recall protocol: the transverse pump ramps exponentially over 8 ms until interactions reach about 2π·2 kHz; |fi| is raised over 3 ms to roughly the recoil frequency ωz ≈ 2π·7.5 kHz, held for another 3 ms, then switched off so the network falls into an unbiased attractor. Holographic imaging of the cavity output reads every spin's phase.
Two ingredients sit outside the Hopfield model. Cavity cooling makes energy-lowering flips more likely in proportion to the energy released, so the path is closer to deterministic steepest descent than to MH, where every downhill flip is equally likely. And J is not frozen. The tweezers are soft enough that optical forces from the cavity field can shift ensembles by up to 4 μm, about 20% of the tweezer waist, so Jij = J(ri, rj) depends on position. A spin flip changes the cavity field, the field moves the atoms, and the current basin deepens, a polaron-like self-trap. The J matrices are not trained to target patterns. The team randomly shifts the array to draw different disorder realizations, then samples natural memories with random stimuli. Five J's are measured at each of n = 4, 8, 12, 16. A pattern counts as a memory only if its basin survives at least one random spin flip at 50% recall probability. At n = 16 the five glasses yield 51 candidates on average; most basins are smaller than one flip and are discarded.
A finite-size Hopfield baseline, recomputed with the same one-flip / 50% rule, stores 3.59 memories on average at n = 16, peaking when P = 4 patterns are intended. The thermodynamic formula 0.138n gives only 2.21. A pseudoinverse learning rule caps capacity at n.
| setting | n=16 capacity | note |
| Hopfield Hebbian | 3.59 | 10,000 disorder averages |
| experiment, five J's | 11.9(6) | 3.3× Hopfield |
| J1, default elasticity | 14(1) | one disorder draw |
| J1, 4× weaker traps | 25(2) | 7× Hopfield |
| experiment at 75% threshold | 5(1) | still above Hopfield |
On J1 the mean basin is 2.1 spin flips versus 3.9 for Hopfield: more memories, shallower wells. Memory 1 reaches nearly four flips; typical memories sit between one and three. Measured capacities track Sherrington-Kirkpatrick simulations under steepest descent and miss the MH curve. Average basin volumes come close to the upper bound 2ⁿ / capacity.
A semiclassical simulation reproduces the counts: 12(2) versus experiment 14(1) at default elasticity with noise, 23(3) versus 25(2) in the soft-trap case. Kill stimulus-phase noise and 0.5 μm trap drift and the same model reaches 36(2) (default) and 60(4) (soft). Without elasticity, trap drift cuts capacity from 22 to 14, about 36%; with enhanced elasticity, trap noise barely moves the number. At n = 20 the team found 97 candidates, tested the 25 most frequent, and counted 11(1) valid memories, a lower bound.
For people building physical neural nets or photonic computers, this is an experimental reversal of "glasses cannot be memories." The capacity gain does not require higher-order couplings. It comes from nonequilibrium descent plus elastic synapses. Short-term synaptic plasticity has long been known to raise biological capacity; atomic motion is a quantum-optical precursor of that mechanism.
It is not a product. Cataloging capacity is already slow at n = 16, and brute-force search fails beyond that. J is not trained, so the stored patterns are whatever the glass grows. The semiclassical model matches capacity but misses the actual patterns by about six spin flips on average, because the analytic J is imprecise. The practical hardware follow-up is phase-locking the longitudinal and transverse pumps and replacing the DMD with a spatial light modulator: the same simulation says killing noise can multiply default-trap capacity by two to three.
The experiment never records the spin trajectory, so "steepest descent" is an inference from matching SK-SD capacity curves, not a dynamical measurement. Long-term aging of the glass is unexplored. The 50% recall threshold was chosen to suppress finite-size effects and lab noise; at 75% the n = 16 average falls to 5(1) and the soft-trap bonus shrinks. An unlocked global phase on the stimulus can weaken |fi| and hurt recall. The n = 20 network is only partially mapped.
Turning spurious minima into memories buys shallower basins, so noise tolerance is worse than Hopfield. The elasticity is short-term: drop the pump and the atoms spring back to the tweezer centers. Lasting plasticity is still a sketch, for example extra minima in the optical potential that freeze learned positions.