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

Abstract The escalating energy demands for high-performance machine learning have sparked growing interest in unconventional computing paradigms rooted in physical systems. At the core of this emerging direction is the leveraging of physical properties to design computing frameworks that integrate strong learning capabilities, efficient training strategies, and practical implementability. Here, we present a simple, efficient, and scalable framework aimed at achieving this, and validate it on an optoelectronic platform. This framework features our proposed training mechanism—state-skipping direct feedback alignment—which eliminates access to intermediate states of the system and thereby significantly simplifies the error backpropagation process, enhancing both training efficiency and practical feasibility. Compared to conventional deep neural networks, our approach substantially reduces computational costs and training parameter count while achieving comparable performance. Furthermore, we integrate our scheme into modern architectures, attaining improved performance with an approximately 40% reduction in computational resources. Notably, our approach challenges the scaling laws of conventional counterparts, underscoring its strong scalability and practical promise.