Decoding Neural Information Of Pigeon Choice Behavior Based On Few-shot Learning

Neural information decoding is a key technology that reveals the internal information processing mechanism of the brain by quantifying the relationship between external behavior representations and neuroelectric activity information,which is of great significance to the development of brain-computer interfaces.However,due to the instability of the internal cognitive and bahavioral states of subjects and the difficulty of ensuring the reliability of the signal quality after electrode implantation,the size of effective sample in the experiment is small.Therefore,how to use small size samples to improve the performance of neural information decoding is important for implantable brain-computer interfaces.Aiming at the above problems,this paper chooses pigeons as the model animal because of their good decision ability and the Nidopallium Caudolaterale(NCL)relates to cognitive as the target brain area.Firstly,an experimental paradigm of pigeon operant conditioning was designed based on value choice and then the neural signals of the NCL area under the decision-making task were collected.Finally,this paper based on transfer learning,combined with the method of brain function network and multivariate Gaussian model,explores the method of decoding pigeon's decision behavior with small size samples.The completed work is summarized as follows:(1)The experimental paradigm of pigeon operational conditioning based on value choice was designed,and the synchronous collection of behavioral data and neural signals of pigeon choice during the experiment was completed.Firstly,this paper designed an experimental paradigm for pigeons to choose the condition of red and green light stimulation with food as a reward.Secondly,we simultaneously collected the 16 channels of neural signals in the NCL brain area of the pigeons and behavioral data during the value choice learning process.Finally,combining behavioral data to determine the reaction time of the pigeon's decision behavior,and through the analysis of the correct rate of the pecking key,the pigeon learning process is divided into two stages: exploration and acquisition.(2)The local functional network of neural signals in the NCL area in the pigeon's value decision process was constructed,and the network topology features were analyzed,and then the separability features of different decision behaviors in different learning stages of the pigeon were extracted.Specifically,we first determined the response frequency band of the local field potential signal when the pigeon is making a decision using the power spectral density method.Secondly,based on the complex network theory,we constructed the coherence local functional network of the NCL area of the pigeon in different learning stages and different decision behaviors from the perspective of multi-channel correlation.Finally,we extracted the two separable features of clustering coefficient and global efficiency based on the topological features of the functional network.(3)A few-shot learning method based on a multivariate gaussian model is proposed to realize the neural information decoding of pigeon's decision behavior under small size samples.Firstly,we established a source domain model for the data of any two pigeons according to the extracted functional network feature distribution.Secondly,based on the idea of transfer learning,we take the data of the remaining pigeon as the target domain,use a small amount of samples in the target domain to adjust the source domain model to form a new model,and use it to decode the behavioral neural information of pigeons in the target domain.Finally,we compared the proposed algorithm with other traditional machine learning decoding algorithms,and the results show the advantage of the algorithm in this paper of high decoding accuracy under small size samples.