Decoding Of Pigeon Movement Direction Based On Partial Least Squares Features

Brain-Machine Interface is the interface between brain and external device, which converts neural signals to control signal by recording and decoding neural signals into motion intent. The Brain-Machine Interface has a great significance to society and is of strong applied value in disabled persons assistance、neurologic diseases treatment and cognition research. Neural decoding is critical in accurate real-time control of Brain-Machine Interface. Because of the noise from the animal itself and external environment,and the sparse speciality of brain signal coding, the neural ensemble signals recorded by microelectrode contain a lot of noise and redundant information, which increase the complexity of model learning, make parameter estimated invalid and reduce the model stability,and eventually make it difficult to realize real time control of external device. To solve this problem, the thesis explores the feature extraction and decoding methods of neural ensemble signals based on pigeon nidopallium caudolaterale signal. First, the thesis uses effective feature extraction method to remove the large amount of noise and redundant information, then an appropriate classification model is choosed to increase the accuracy and stability of neural ensemble decoding.First, the relativity among the neural signals recorded by each channel during pigeon moving in the plus maze is analyzed, the results indicate that because of synchronous firing of neurons, signals between channels contain large amount of redundancy. At the same time, due to the interference from the external environment and the pigeon itself, some channels contain large amount of noise irrelevant to decoding objective. So the feature extraction is very necessary before neural decoding. The feature extraction results of neural ensemble signals by partial least squares method indicate that the movement directions can be distinguished with only three latent variables. Compared with principal component analysis method, partial least squares extract fewer features with more useful information related to decoding objective. Therefore, the partial least squares method can be used as an effective feature extraction method in decoding of pigeon movement direction.For the features extracted by partial least squares method, three classification models including support vector machines 、 K-nearest neighbors and linear discriminant analysis are used to distinguish different motion intents. The results show that among four models, the linear support vector machines perform best in decoding accuracy、stability and efficiency. The nonlinear support vector machines has the problem of overfitting and parameter selection. K-nearest neighbors also has high decoding accuracy, but it is not as good as the linear support vector machines in stability and efficiency. So the best classification model based on the partial least squares features in decoding of neural ensemble signals is linear support vector machines.