In an era marked by the rapid development of artificial intelligence neural networks, traditional electronic processors are under immense strain. This is due to the large - scale matrix operations and frequent data iterations that these networks demand. Optoelectronic hybrid computing, which ingeniously combines optical and electrical processing methods, exhibits outstanding computational capabilities. Nevertheless, in real - world applications, several challenges arise. The separation of training and inference processes, along with offline weight updates, can lead to a decline in information entropy. This, in turn, results in reduced computational accuracy and low inference accuracy.
