Two Holograms Match Modern Hopfield Networks for Optical AI Memory

Researchers David J. Brady and Gregory Neory have published a paper on arXiv proposing a physical optical implementation of modern Hopfield networks (dense associative memories) using two volume holograms. Associative recall—mapping an input pattern to the stored pattern it most resembles—is a natural operation for a volume hologram, but direct 2D-to-2D holographic implementations suffer from Bragg degeneracy, forcing fractal sampling. The authors show that a cascade of two volume holograms separated by a 1D coded layer physically evaluates the modern Hopfield retrieval map η = V softmax(λK^T x) exactly as a parallel optical computation. The inverse temperature λ is realized via optically addressed spatial light modulation in the coded layer. Routing input and output through a 1D code rather than directly between 2D planes supplies the separating nonlinearity that the original Hopfield model lacked and removes the degeneracy by balancing the grating-wavevector dimension count to 2+1=3.

Faithful dense storage further requires a recording medium that captures inter-neuron connections while rejecting the field self-energy responsible for the M^{-2} efficiency falloff of homogeneous photorefractives. To solve this, the authors propose a nonlocal, gradient-responsive medium whose illumination-independent decay recovers linear M^{-1} scaling in situ. They demonstrate its reception, combination, and storage functions in a discrete opposing-diode cell. The paper also outlines routes to OASLM-stack and volume molecular/nanocrystal realizations. This work bridges nonlinear optical computing with modern deep learning, potentially enabling ultra-fast, high-capacity associative memory for vision systems that operate at the speed of light.