Brief

Researchers have developed a deep photonic neuromorphic network (PNN) architecture that utilizes phase-change material (PCM) synapses and local optical feedback for unsupervised Hebbian learning. This novel approach bypasses the need for external gradients or complex electro-optical conversions by directly employing correlated pre- and post-synaptic optical activity for adaptation. Experiments using fiber-optic components and programmable attenuators demonstrated the system's ability to achieve adaptive synaptic evolution, optical inference, and autonomous pattern encoding, paving the way for energy-efficient, integrated photonic neuromorphic systems. AI

Researchers have benchmarked seven different Hebbian learning rules for their effectiveness in associative memory tasks, specifically focusing on prototype extraction. The study evaluated pattern storage capacity, information capacity, and the ability to recall correct prototypes from distorted instances using recurrent networks. Bayesian-Hebbian learning rules demonstrated the highest capacity across various conditions, outperforming simpler additive Hebb rules and covariance learning. AI