The human brain exhibits remarkable capabilities in parallel perception, reasoning, and learning, all while consuming a mere 20 watts of power. This efficiency serves as a profound source of inspiration for the advancement of artificial intelligence (AI). At present, the majority of AI hardware still relies on the von Neumann architecture, which necessitates frequent data transfers between memory and computing units. This constraint leads to limitations in bandwidth, energy efficiency, and scalability. Optical computing, distinguished by its high-speed data propagation, inherent parallelism, and high-dimensional interconnection capabilities, presents a promising avenue to overcome these bottlenecks. Nevertheless, existing optical neural networks continue to depend on external optoelectronic conversion and digital buffers for interlayer connections and weight reconfiguration, thereby impeding the progress towards large-scale, deep optical neural networks.