Deep convolutional neural network system based on laser array and control method

Abstract

The invention relates to a deep convolutional neural network system based on a laser array. A deep convolutional layer comprises a plurality of transversely coupled semiconductor lasers, a plurality of optical couplers and a plurality of feedback loops; an input optical signal is injected into the first semiconductor laser, a signal output by the nth semiconductor laser is divided into two paths through the nth optical coupler, and one path of signal is returned through the nth feedback loop and serves as an input signal of the (n + 1) th semiconductor laser; the other path of signal is input to the output layer; the output layer comprises an adder and a programmable gating array; and the other path of signals output by each optical coupler are superposed in the summator to form output information, and the programmable gate array trains the output information to obtain the weight of the output layer. A plurality of transversely coupled semiconductor lasers are arranged, so that the training of interlayer connection can be avoided; and only the output of the output layer needs to be trained, so that the dependence on an algorithm is greatly reduced, and the difficulty in hardware implementation is also remarkably reduced.

Description

technical field [0001] The invention relates to the technical field of machine learning, in particular to a deep convolutional neural network system and a control method based on a laser array. Background technique [0002] Artificial neural networks have become the current disruptive computing concept. When multiple network layers are cascaded, these systems achieve great success in multiple challenging tasks. In such deep neural networks, different layers are used to emphasize specific aspects of the output information, while the input of each network layer is also used as the output signal of the next layer. This continuous hierarchy is of great significance for improving computing performance. In the past decade, deep convolutional neural networks (CNNs) have achieved remarkable achievements in different fields, including still image recognition, etc. [0003] In a deep convolutional neural network (CNN), each layer convolves its input with a spatial filter. By incre...

AI Technical Summary

Problems solved by technology

[0004]However, the current deep convolutional neural network needs to be trained for inter-layer connections, which has high requirements for algorithms and hardware, and it is difficult to directly map the network topology in the future to the physical substrate, resulting in slower information processing rates and significant challenges in hardware integration

Images