Photonic neuromorphic learning via generalized in situ physical gradient descent.

INSPIRE achieves in situ photonic neural network training with a relative error of 0.26%, enabling efficient, adaptive optical learning systems.

• Why it matters: Current physics-based neuromorphic computing faces challenges in training due to costly modeling and fabrication errors, limiting practical deployment of photonic systems.
• What they did: The authors developed INSPIRE, a general on-chip training method utilizing synthetic time-reversal holography to measure complex photonic modes and perform gradient updates directly in physical circuits, demonstrating its effectiveness on matrices larger than native tunable elements.
• The result: This approach enables single-shot photonic learning with 251-fold model compression and 136-fold acceleration, paving the way for practical, adaptive, and energy-efficient intelligent photonic devices.