Investigation On Improving The Information Processing Rate Of A Reservoir Computing System Based On A VCSEL Under Variable Polarization Light Injection

Reservoir computing(RC)that emerged in the 21 st century,is a kind of recurrent neural networks(RNNs)algorithm.It can be divided into spatial RC and time-delay RC,while time-delay RC based on a nonlinear node has a simpler structure and lower hardware implementation cost.Due to the unique advantages of semiconductor laser(SL),such as fast response,low energy consumption,and rich and controllable nonlinear dynamic states when subjected to external disturbances,SL is very suitable as a nonlinear node for time-delay RC.As a typical type of SL,vertical cavity surface emitting laser(VCSEL)also has the advantages of small size,low threshold current density,and easy integration.In particular,under appropriate operating parameters,it can simultaneously provide two mutually perpendicular polarization components(X-PC and Y-PC).Therefore,due to its unique luminescence mechanism,many research results on time-delay RC based on VCSEL have been reported.In the currently proposed VCSEL-based RC system,the polarization direction of injection light loaded by the information is along either X-PC or Y-PC of the VCSEL,which is named as conventional information injection(CII).In this case,a polarization component is associated with only one processing task.Therefore,there is still more exploration space for the two PCs of VCSEL.In this thesis,we propose that the injection light loaded by the information is set at45-degree from the X-PC(named as variable polarization information injection,VPII)to improving the information processing rate of the VCSEL based RC system.Under this case,the information is simultaneously injected into two PCs,and therefore the virtual node states extracted from X-PC and Y-PC can be combined for training and testing so as to the number of virtual nodes can be doubled.Compared to the CII scheme,the VPII scheme can double the information processing rate for a given interval of θ through reducing the number of virtual nodes by half,and its effectiveness is numerically verified via waveform recognition task(WTASK)and Santa-Fe chaotic time series prediction task(STASK).To double the processing rate under the VPII scheme,the number of nodes for each PC is set to half of the number under the CII scheme.Two different scenarios are considered.One is for the VCSEL-based RCs are utilized to process single task,and the other is for the VCSEL-based RCs are utilized to process two tasks in parallel.We first discussed the effects of feedback strength kd and injection strength kinj on the Normalized Mean Square Error(NMSE)value for single task processing,where polarization-preserved optical feedback(PP-OF)and polarization-rotated optical feedback(PR-OF)are considered in each benchmark task.For Santa-Fe time series prediction task,the kinj-kd parameter regions of NMSE ≤ 0.1 for VPII scheme are consistent with the regions for CII scheme or even wider,and the minimum NMSEs with VPII scheme are smaller than the values with CII scheme.In addition,waveform recognition task is also discussed on the kinj-kd parameter regions.Next,the system performance for processing two tasks in parallel is also discussed on the kinj-kd parameter regions.Two injection lights,loaded by WTASK and STASK respectively,are injected into VCSEL at the same time via adopting VPII scheme,and both X-PC and Y-PC contain the information of two tasks.The results indicate that,for both STASK and WTASK under two types of feedback,the kinj-kd parameter regions of NMSE ≤ 0.1 for VPII scheme are consistent with the regions for CII scheme or even wider.The results show that whether processing one task or two tasks in parallel,after adopting VPII to replace CII,the processed rate can be doubled and meanwhile the performance of the RC is comparable to that obtained under adopting CII.In following research,we optimize the parameters of the parallel RC system adopting VPII scheme to obtain the minimum NMSEs for two tasks.