Computational Error Analysis And Design Of Optical Computing

In the information age,the rapid expansion of data promotes the rapid development of artificial intelligence technology,but also brings heavy pressure to the hardware computing platform.Although electronic hardware has been redesigned from different angles to speed up deep learning computing,it still cannot solve the dilemma that Moore’s law is facing failure.At the same time,optical computing,which has the advantages of passive computing,low latency,low loss and massive parallelism,has entered people’s field of vision,and the optical computing circuits based on this are also expected to become the key to breaking the situation.Recent years,optical computing circuits develop rapidly,and designs based on different optical components or different optical principles have sprung up,making people see the hope of replacing electronic computing with optical computing.However,the noise in components accumulates continuously,and the computational error caused by it makes the expansion of optical computing circuits become a difficult problem.What kind of influence will the error causes to the circuit-based networks,how to control the error,the solution of these problems will greatly promote the development of optical computing circuits.This paper mainly studies the computational error in optical computing circuits.The main work content and innovations are as follows:(1)Based on the problem of noise accumulation in optical computing circuits,this paper establishes a computational error model,and obtains the error growth trend through simulation calculations.The error model starts from the two dimensions of absolute distance and relative distance,takes the matrix dimension,noise intensity and parameters themselves as variables,which makes it possible to describe the error curve more comprehensively and conduct systematic analysis on the error.In addition,the error model introduces the representation of state transition diagram,which makes it could analyze specific components more accurately while reducing the amount of simulation calculations.During the experiment,this paper found that the absolute distance average error under certain circumstance will show a certain convergence trend,which brings a transformable error observation method and also brings new ideas to error control.However,the overall experimental results show that the continuous and regular growth of error will seriously limit the integration and scale development of optical computing circuits.(2)Based on the problem that optical computing circuits are difficult to be expanded and integrated on a large scale due to noise,this paper proposes a network structure design with higher noise tolerance.This structure uses the single component as a neuron,which improves the utilization efficiency of components and reduces the depth of the network.In addition,the network design breaks away from the previous design ideas,starting from the neurons represented by components rather than from the weight matrix,so it is more flexible and has more possibilities for expansion.The experimental results show that the design can reduce the usage of components by an order of magnitude and improve the anti-noise ability of the network without losing too much accuracy.