Research On Performance Optimization Of Delay-Based Reservoir Computing Using Optical Injection

In the last few years,great progress has been achieved in artificial intelligence technologies such as neural network algorithms and related applications.However,hardware implementations of these algorithms face a lot of challenges due to high loss and large delay of electronic neural networks.The reservoir computer(RC)based on nonlinear optical devices has been proposed to compensate these drawbacks.The nonlinear characteristics of optical devices are utilized in RC to achieve the function of hidden layer in neural network algorithm and the training process only exists in the output layer.Compared to echo-state-networks,the delay-based RC that consists of a nonlinear node and a delay loop simplifies the structure.The RC has been successfully implemented in various tasks such as speech/handwritten digit recognition,nonlinear channel equalization,optical packet head recognition,and time-series prediction.As a result,the work thesis focuses on the optimization of delay-based RC where a semiconductor laser is adopted as the nonlinear node.Based on the characteristics of optical injection and feedback,a conventional all-optical delay-laser-based RC is analyzed and tested.The task of chaotic time-series prediction is used to evaluate the performance.The injection locking state is necessary for the task obtained by setting the negative detuning frequency and injection coefficient of injection laser.Using the conventional RC with a long delay time,the NMSE of time-series prediction is calculated to be about 0.0028.However,the performance degrades for using shorter delay time.The result indicates that the conventional RC cannot maintain good prediction performance in high processing speed.Towards the issues above,the motivations of the thesis are to increase the data processing speed and optimize the neural training method.A delay-laser-based RC usually has its processing speed limited by the transient response of laser dynamics.Here,we study a simple all-optical approach by introducing optical injection to the reservoir layer of conventional RC.By using optical injection,the laser’s transient response time effectively decreases thus increases the processing speed.In the chaotic time-series prediction task,the proposed RC achieves good performance in a flexible range of injection detuning frequency under sufficient injection rate.The prediction error is significantly reduced and stabilized at high processing speed.For achieving prediction error below 0.006,the optical injection enhances the processing speed by about an order of magnitude to 5 GSample/s.As for the method of neural training,the performance of a conventional RC usually deteriorates dramatically if parallel interpolation is adopted for data serialization in the input layer.In order to solve the training problem,the many-to-one for input-output mapping method is adopted to replace the one-to-one method.The results show that the performance of timeseries prediction is improved in same circumstance and the robustness of the reservoir computer is enhanced.Optimal performance is achieved when the number of input data is 4.In addition,by reducing the amount of data training,the proposed method reduces the cost in terms of running time and data space while maintain comparable performance.