Chip de red neuronal óptica coherente con un nuevo modelo de computación para multiplicación de matrices-vectores a...

In this work, an innovative optical neural network (ONN) chip based on the coherent detection structure has been proposed. By establishing a mathematical model, the approach eliminates the need for phase compensation while reducing the encoding and decoding process by integrating the transformation function into the activation function. This improves the computing efficiency while maintaining high accuracy. The performance of the chip has been evaluated across multiple image classification datasets, with its accuracy on the MNIST dataset reaching 97.28 %, comparable to the results run by a computer. Furthermore, the design demonstrates strong robustness and generalizability accommodating diverse network architectures ranging from single-hidden-layer networks to more complex multi-hidden-layer configurations.