Zero-power optical convolutional neural network using incoherent light

Abstract

As a new high-speed intelligent computing method, optical neural network (ONN) has the advantage of realizing lower or even zero power computing. For example, the diffractive neural networks based on passive diffraction layer can work without power consumption. However, in the realization methods of current ONN, coherent light is used to carry input information, while the light in the natural scene is incoherent light, which makes it difficult to apply ONN to the actual optical scene for physical observation. In this paper, we propose an irregular incoherent optical convolutional neural network (I2OCNN). The network only uses the reflection and transmission properties of light to realize the controllable rearrangement of two-dimensional incoherent light field on a series of passive optical devices, so as to realize the cross interconnection of optical neurons, and thus realizes the optical convolution neural network of incoherent light with zero power consumption. Since the architecture is based on intensity modulation and incoherent superposition, incoherent light can be used to carry input signals, which solves the problem of practical application of ONN. In addition, the network can achieve irregular convolution. MNIST and Fashion MNIST were used to verify the image recognition capability of 7-layer I2OCNN, and the testing accuracy was 86.95% and 75.68%, respectively. Theoretical reasoning and simulation results show that this architecture can complete the basic image recognition tasks under the condition of incoherence and zero power consumption.