基于非监督学习的互适应脑机接口神经信息解析

项目名称： 基于非监督学习的互适应脑机接口神经信息解析

项目编号： No.61473261

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 其他

项目作者： 王怡雯

作者单位： 浙江大学

项目金额： 83万元

中文摘要： 由于无法获得丧失运动能力的患者实际的肢体运动轨迹，脑机接口的临床应用面临挑战：1）不能使用传统监督式学习的方法解码；2）脑必须适应外部设备产生的运动行为，实时调整锋电位发放的模态（可塑性），智能的外部设备需要及时跟踪神经元的动态变化。因此，脑机接口本质上是一种基于非监督学习的互适应在线交互系统。本项目以非人灵长类动物（猴）为主要实验对象，针对脑的非稳态神经元活动，研究非线性可视化分析与建模方法，预期变化规律，为解码提供先验知识；针对智能外设，研究基于无监督或半监督式学习的锋电位动态解析算法，通过扩展双模型的贝叶斯动态估计和研究连续状态-行为空间的强化学习解码方法，实现连续运动的神经解析，提高脑机接口性能，延长使用时间；针对在线交互需求，建立基于数据驱动的神经元评估方法，动态构建重要神经元子集，减少解析计算量，提高脑机交互的计算效率；设计互适应实验平台，基于神经控制实现连续运动及运动规划。

中文关键词： 脑机接口；锋电位解码；互适应；非监督学习；植入式

英文摘要： Brain machine interfaces (BMI) cannot directly obtain the actual movement of the disabled patients in clinical application，which brings challenges to the traditional decoding methods by supervised learning that trains by the error between the model outputs and desired movements. The brain needs to adapt to the behavior of the external devices and generates new firing patterns due to its plasticity, while the intelligent external device needs to simultaneously tracking the non-stationary neural activities during decoding, which forms a co-adaptive system. In this proposal, we build an invasive co-adaptive BMI platform on primates and study the time-variant properties of neural activities and dynamic kinematics, and the online interaction between brain and machines. We propose to visualize the high-dimensional neural activities in feature space and parametrize the dynamic tuning properties, decode the non-stationary neural firings by unsupervised and semi-supervised learning methods, including a dynamic kinematic state estimation by Bayesian approach using dual model, and reinforcement leaning algorithm for large state-action space to realize neural control on continuous movement. We also propose a data-driven method to evaluate important neuron subset by analyzing the local structure of the spiking data. Dynamically updating the membership of the neuron subset could reduce the computational burden for online interaction of the co-adaptive BMI. We will testify the above methods for the co-adaptive BMI system by a neural control task with upper limb continuous movement and motion planning.

英文关键词： Brain machine interfaces;spike decoding;co-adaptive;unsupervised learning;invasive BMI