Unsupervised online learning in photonic neural networks

We present a fully optical deep photonic neuromorphic network (DPNN) that supports unsupervised online learning using a local Hebbian learning rule. Unlike conventional photonic systems that rely on offline training and energy-intensive optical-electrical-optical (O-E-O) conversions, our architecture operates entirely in the optical domain. In a tabletop proof-of-concept using fiber-optic components, the network achieves 100% accuracy on a letter recognition task under both supervised and unsupervised learning. We also explore the use of phase-change materials (PCMs) as promising candidates for enabling non-volatile optical synapses and neurons. While not yet implemented on-chip, integrating PCM-based devices is a key future goal to enhance scalability and efficiency. This work establishes a pathway toward next-generation energy-efficient, scalable photonic AI hardware.