Scalable physical deep learning using optical dynamics with state-skipping training and single-map architecture.

A physical deep-learning framework on an optoelectronic platform uses state-skipping direct feedback alignment and a single-map 1D dynamics architecture to cut computational resources by ~40% while matching performance.

• Why it matters: Conventional deep learning faces rising energy and compute demands, and physical computing needs methods that simplify training and reduce complexity without sacrificing learning quality.
• What they did: The authors train with state-skipping direct feedback alignment, avoiding access to intermediate system states, and use a single-map architecture built from simple one-dimensional dynamics.
• The result: Compared with conventional deep neural networks, the method substantially reduces computational cost and trainable parameter count and, when integrated into modern architectures, improves performance with an approximately 40% reduction in computational resources.

AI summary of the abstract · may contain errors · not medical advice · read the paper.