脑机接口是连通大脑与外界的信息通路，它可以用来控制外部设备（神经义肢）或将外界信息传导到大脑当中（人工耳蜗）。按照采集神经活动的信号的方式，它可以分为非侵入（脑电图）、半侵入（皮层脑电）和侵入式（植入电极）三种。控制外部设备的侵入式脑机接口中，主流研究为运动皮层脑机接口。它可按照范式不同分为两类：仿生式脑机接口和生物反馈式脑机接口。仿生学脑机接口使用较为复杂的算法用计算机模拟从神经活动到运动信息的前向通路，通过运用多种解码算法可以达到较好的控制效果；生物反馈式脑机接口则是运用相对简单的算法把大脑和机器连接，让大脑自我适应。前人在这方面的工作数量有限：之前的的研究证实了在啮齿类和灵长类动物中，使用至多四个神经元控制一个维度的光标或物体。我们认为，通过使用生物反馈式脑机接口，非人灵长类可以用多神经元控制多维度光标或物体。 我们在猕猴的初级运动皮层植入电极阵列，采集信号并用这个信号作为脑机接口算法的输入，同时将输出的结果显示在屏幕上。我们构思并实现了多种脑机接口算法参数以及任务策略，选取用其中最优的组合让猕猴进行长期的学习，即使用这样的脑机接口做任务。由于脑机接口的算法是人为规定的，猕猴在刚开始的任务表现表现不佳，然而经过较长时间的训练，猕猴的任务表现提升。我们对猕猴在学习脑机接口过程中的数据进行了分析，得出了以下结论： （1）两只猕猴都有学习脑机接口的表现，在单位时间内完成的试次数量增多，且完成率上升。我们对猕猴完成任务的光标轨迹和神经元群组放电率分析，结果同样支持猕猴学会了生物反馈式脑机接口。 （2）对采集到的单个神经元的调谐曲线进行分析和对多个神经元进行的同一事件（Unitary Events）分析并没有提供足够证据支持猕猴学会此类脑机接口，我们还需要进一步改进分析方法。

BMIs (Brain-machine interfaces, or BCI, brain-computer interfaces) are information pass way witch could control devices, like neural prosthesis, or convert information into brain, like artificial hearing prosthesis. In controlling BMIs, the mainstream researches focus on BMI on motor cortices. It contains two parts as they have different paradigms: bio-mimetic and biofeedback. The bio-mimetic BMIs employ relatively complex algorithms to mimic the information forward pass way from neural activity to motion information. It improve its control quality by using different decode algorithms; The biofeedback BMIs, on the contrary, using relatively simple algorithms connecting brain with machine, and let the brain to learn by itself. The number of works on this kind of BMI is limited: former researches prove that it is possible to let rodents and non-human primates to use up to four neurons to control one dimension cursor or objects. We think that, non-human primates could employ biofeedback BMI, using multi-neurons to control multi dimension cursor or objects. We implanted an electrode array in the primary motor cortex of macaque monkeys, collected the signal, and used it as input to the BMI algorithm, while displaying the output on the screen. We have conceived and implemented a variety of BMI algorithm settings and task strategies, and selected the best for the macaques to learn for a long time, using such a brain-computer interface to do the task. Since the algorithm of the BMI is artificially specified, the macaques did not perform well at the beginning, but after a long period of training, the macaques learned to use such a brain-computer interface. At the same time, we analyzed the data of macaques in the process of learning brain-computer interface, and reached the following conclusions: (1) Both monkeys have the performance of learning brain-computer interface, the number of trials completed in unit time increases, and the completion rate increases. We performed analysis on cursor trajectories when monkeys finished the tasks and the firing rate of neuron ensembles. The results support the monkeys have learned the bio-feedback BMI. (2) The analysis on tuning properties of single neurons and the Unitary Events on neuron esembles does not provide enough evidences that monkey have learnt this kind of BMI. Thus, we need to improve our analysis methods to find more evidence to support that monkey have learned this kind of BMI.