Time Correlation And Weakly Supervised Learning In Neural Decoding For Macaque's Moving Finger

By studying the neural circuit's perception of the external world and producing behavior,it can enhance the ability to sense and control the outside world.In the process of neural information processing,the encoding and decoding of neural signals is an important part in this research.In this paper,a healthy macaque is selected as the research object,and the signal of the peak potential of the macaque motor cortex neurons collected by the neural electrode array is used to estimate the position of the finger movement.This process is simply referred to as neural decoding.Of course,the neural coding process can be understood as the "reverse process" of the above neural decoding.Therefore,the neural coding is the acquisition of brain neural signals by means of external devices.Next,the decoding of neural information processing will be studied and discussed from the following aspects:1.The time correlation of the traditional TILM model is analyzed.A Convolution Space Model(CSM)is derived from the State Space Model(SSM)to decode the position of the macaque's moving finger.The model still belongs to the TILM model,but compared with the traditional model,the CSM can not only correlate the current moment state with the previous moment,but also the state of the previous multiple moments.The experimental results show that the decoding error of the traditional method is greater than that of the CSM model.2.The original SSM model is improved by combining the CSM model.In the observation equation,the correlation is extended from the peak potential signal at the previous moment to the peak potential signal at the multiple previous moments.The CSM model is implemented using a convolutional network for decoding.In the study of weak supervised learning,it is found that there are fewer sample points collected at the boundary of the moving range,and the decoding result is too large to be eliminated.The experimental results show that the CSM model and the improved SSM model have better decoding accuracy than the traditional model.3.Supervised decoding methods are gradually becoming mature,and the unsupervised processing of regression problems is too large at error.In order to solve this problem,Weakly Supervised Learning(WSL)was used to correct the inaccuracy of the unsupervised training finger position and further fit the accurate weight parameters.The experimental results show that the error of the WSL decoding position is close to supervised on the two-dimensional plane,with a difference of only about 0.4%,and compared with the unsupervised decoding error,with an increase of about 41.3%.4.A generative adversarial network(GAN)is introduced,which generates a large number of training data with the same distribution as the original data,and uses the WSL model to decode the finger movement position.The experimental results show that the WSL model with GAN has higher accuracy.