生物脑控机器人的研究和开发
文献类型：学位论文
| 作者 | 李永程
|
| 学位类别 | 博士 |
| 答辩日期 | 2015-11-25 |
| 授予单位 | 中国科学院沈阳自动化研究所 |
| 授予地点 | 中国科学院沈阳自动化研究所 |
| 导师 | 王越超 ; 李洪谊 |
| 关键词 | 神经-机器人混合系统 人工智能 智能机器人 神经网络 神经接口 |
| 其他题名 | Investigation and development on biological brain robots |
| 学位专业 | 模式识别与智能系统 |
| 中文摘要 | 本文在大量的国内外文献调研和综合的基础上，采用胎鼠海马细胞培养的离体神经网络作为生物控制器，针对上述问题，开发了一个完整的生物脑控机器人系统，并且在这个系统上开展了基础实验研究和工程应用研究，具体的内容如下：(1) 根据Hebbian 理论，我们进一步研究了细胞集群活动的时空动态性。我们在多电极阵列上培养了胎鼠的海马神经元并利用单一电极通过电刺激诱发神经网络的回响活动。然后，我们分析从多电极阵列上记录到的神经网络回响实验数据集，并得知：在网络展示的诱发的回响中，针对每个特定的刺激输入，在神经网络的活动中通常都包含一个精度为几毫秒在整个回响发放期间重复很多次的主时空发放模式。有趣的是，在同一个神经网络中，不同的输入点所诱发的时空模式是不同的，尽管有时他们可以共享重叠的亚神经元网络群，并可能会形成一个核心电路。这些结果说明，具有精确的时空发放的神经网络活动模式在神经网路回响中能够被神经网络有效的维持着，这种方式可能是神经网络进行精确和有效的信息编码的方法。(2) 为了研究神经网络的时空动态性，我们开发了一个有效的策略在电刺激诱发的神经活动中来找到保守的重复时空模式并且评估了它们的统计有效性。通过利用高维特征空间中的相似性函数，我们可以确定，在许多个回响事件中（即在几十毫秒尺度下同时涉及多个神经元发放的一种神经网络活动），显示出了精度为几毫秒的相同的时空模式。同时通过算法我们还可以确定，在相同神经网络中，不同的电刺激所诱发的这些时空模式是几乎不一样的。这些发现再次表明，神经回路能够保持回响活动的精确的时空模式，这也意味着寻找这些时空模式的方式可以作为一种高效的信息解码方式应用于生物脑控机器人中。(3) 为了验证层级化的神经网络对生物脑控机器人性能的提高，我们在搭建好的生物脑控机器人系统中采用了两种离体神经网络——‘4Q’ 和随机神经网络作为神经控制器。所谓‘4Q’ 神经网络就是人为的将神经网路分成四个内部互相连通的部分，从而培养神经网络结构上形成层级分化。通过我们的研究，与随机神经网络相比，‘4Q’ 神经网络有着完全不同的神经活动。同时，在目标搜索任务中，‘4Q’ 神经网络控制的机器人表现出了更好的性能。我们的研究结果表明，两种神经网络都可以成功地用来控制人工代理;因为神经网络的短期可塑性，随着实验中外界刺激的持续进行，机器人性能会越来越好，且‘4Q’ 神经网络控制的机器人在这个变化中的性能表现比随机神经网络明显要好的多。(4) 最后，我们建立了自己的整合了机械智能和生物智能的生物脑控机器人系统。在我们初步建立的这个框架中，我们将离体的神经网络和移动机器人系统相连接，实现了一个新颖的车辆设备。移动机器人系统由摄像头和两轮机器人组成并被设计用来执行目标搜索任务。根据开源软件我们在linux 操作系统下开发了一个实时神经信号处理软件，同时开发了一个实时的刺激生成器。这两者可以保证生物和人工部分之间通过简单的二项编码/解码方式实时的进行双向信息交换。由于我们对这个系统和实验的创新研究，一些突出的结果是值得关注的。作为神经控制器，在我们的实验中，我们使用了一个特殊的比在以往研究中的神经网络都复杂的层级神经网络。它表现出了极大的优势，极大的改进了机器人的性能。基于我们的工作，‘4Q’ 神经网络可以以一个较高的性能成功地控制人工代理（单次任务中成功转向率可以到达100%）。一个由于有效的实验方案带来的令人惊讶的发现是，强化刺激训练的情况下，机器人在实验的训练期间性能表现的越来越好，这说明我们已成功的将神经网络的短期可塑性发展阶段应用在了生物脑控机器人中。基于这些结果，我们最终改善了生物脑控机器人系统的性能（在两个方向上的正确转向率平均值超过80%，其机会水平为33%）。这个新框架将为基于生物- 人工双向接口的新型智能机器人系统提供一个可能的实现方案。本文的研究可以为生物脑控机器人的深入研究和开发做进一步的贡献。在体现神经科学和机器人学深度交叉的基础上，我们将利用生物脑控机器人进一步揭示神经编码和神经网络学习的基本规律，同时，将发现的基本规律有效的整合在机器人系统中尽快实现基本规律的工程应用，以期在生物脑控机器人上验证和应用的基本神经规律可以积极有效的应用于神经康复机器人，神经假肢，脑控机器人等实际的医疗和应用系统中。 |
| 英文摘要 | According to investigation on a plenty of papers, we adopt the dissociated neural network cultured from embryonic rats hippocampus to play as the neural controller. To the issues mentioned above, we develop an integrated ‘Biological brain’ robotics system, and do the foundational science experiments and engineering application researches in this system, details include: (1) Based on the Hebbian plasticity theories, we further studied the spatiotemporal dynamics of such collective neuronal activity, we grew hippocampal neurons on multi-electrode arrays (MEAs) and evoked network reverberation via electrical stimuli from single electrode. After that, we analyzed experimental data sets of network reverberation from the MEA recordings and knew that: In networks exhibiting evoked reverberation, the activity following each stimulus of a specific input site often contained a dominating motif with precision of a few milliseconds that repeated many times throughout the whole reverberation period. Interestingly, reverberation evoked at different input sites in the same network contained motifs of distinct patterns, although sometimes they could share over- lapping subpopulations of neurons in the network that might have formed a core circuit. These results demonstrate that activity motifs with precise spatiotemporal patterns could be maintained within a neuronal network during reverberation, and could potentially serve as a precise and efficient way of information coding. (2) In order to study the spatiotemporal dynamics of the network, we developed a useful method to find the conservative repeated spatiotemporal motifs and estimate their significance in the network given an electrical stimulus. According to the usage of similarity function in the high-dimensional feature space, we found that many reverberation events, which synchronously involved many neuron spiking activities in a time window of around tens milliseconds, showed similar firing motifs with precision of a few milliseconds. Given different electrical stimulus to the same network, the motifs are barely the same. These findings suggest again that neural circuits are capable of maintaining precise spatiotemporal patterns of reverberatory activities that could be utilized as an efficient way of information coding. (3) In order to prove that hierarchical neural networks can bring great improvement to the performance of ‘Biological brain’ robots, the two types of experimental preparations were utilized as the neural controller including ‘random’ and ‘4Q’ (cultured neurons artificially divided into four interconnected parts) neural network in the built ‘Biological brain’ robots system. Compared to the random cultures, the ‘4Q’ cultures presented absolutely different activities, and the robot controlled by the ‘4Q’ network presented better capabilities in search tasks. Our results showed that neural cultures could be successfully employed to control an artificial agent; the robot performed better and better with the stimulus because of the short-term plasticity. The robots controlled by the ‘4Q’ neural networks have a significantly better performance than that controlled by the random neural networks. (4) At last, we built a complete ‘Biological brain’ robots system for our experiments. In this framework, we connected the dissociated neural network to a mobile robot system so as to implement a novel vehicle. The mobile robot system characterized by a camera and two-wheel robot was designed to execute the searching task. The modified software architecture and home-made stimulation generator were employed to support a bi-directional exchange of information between the biological and the artificial part by means of simple binomial coding/decoding schemes. Some highlight results caused by our experimental designs for this framework were worthy of attention. Acted as the neural controller, a special hierarchical experimental preparation which was more complicated than previous studies, was utilized in our experiment and showed great advantages to the improvement for the performance of robots. Based on our work, neural cultures could be successfully employed to control an artificial agent at a high performance (in a single trial with a correct turning percentage of 100% ). A surprising find resulted from effective experiment proposal was that, under the condition of tetanus stimulus training, the robot performed better and better with the training period in the experiments because of the short-term plasticity of neural network. We finally improved the performance of the ‘Biological brain’ robots on the basis of these design (the average percentage of the correct turning was over 80% in both directions, chance level 33%). This new framework may provide a possible solution for the development of new strategies for novel intelligent robot system on the basis of biological-artificial bi-directional interface. Our study will give more contribution to the ‘Biological brain’ robots systems for their further investigation. On the basis of the interdisciplinary research between neuroscience and robotics, we will go further step to indicate the basic rules about neural coding and neural learning via using the ‘Biological brain’ robots system, at the same time, effectively employ the discovered basic rules in this hybrid system in order to assess these rules proven and used in ‘Biological brain’ robots system to make them be able to efficiently applied to neural rehabilitation robots, neural prosthesis and brain-controlled robots in future. |
| 语种 | 中文 |
| 公开日期 | 2015-12-27 |
| 产权排序 | 1 |
| 页码 | 101页 |
| 源URL | [http://ir.sia.ac.cn/handle/173321/17533]  |
| 专题 | 沈阳自动化研究所_机器人学研究室 |
| 推荐引用方式 GB/T 7714 | 李永程. 生物脑控机器人的研究和开发[D]. 中国科学院沈阳自动化研究所. 中国科学院沈阳自动化研究所. 2015. |
入库方式： OAI收割
来源：沈阳自动化研究所
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。