Photonic neuromorphic learning via generalized in situ physical gradient descent.

In situ physical gradient descent (INSPIRE) enables highly accurate, real-time training of photonic circuits with less than 0.3% error, advancing adaptive photonic systems.

• Why it matters: Current neuromorphic computing relies on costly, error-prone physical modeling, limiting efficiency and scalability. Developing practical, on-chip training methods is essential for real-world photonic intelligence.
• What they did: The authors developed INSPIRE, a general, topology-agnostic training approach that uses on-chip holography to measure complex photonic modes and perform gradient updates directly within the physical system, demonstrating successful training of matrices larger than native elements.
• The result: This method achieves fast, accurate in situ learning, including meta-learning with significant model compression and task acceleration, paving the way for efficient, adaptive photonic neural networks.