Optical Implementation of Equilibrium Propagation Using Spatial Photonic Ising Machines
Researchers have developed a hybrid optical-digital system to implement Equilibrium Propagation (EP), a machine learning training method for energy-based networks. This system utilizes a Spatial Photonic Ising Machine (SPIM) to optically encode continuous neuron states and trainable patterns. The approach was tested on a wine classification dataset and numerically evaluated on the MNIST dataset, demonstrating a path towards energy-efficient physical implementations of EP. AI
IMPACT Demonstrates a novel hardware approach for training energy-based models, potentially leading to more energy-efficient AI systems.