基于迁移学习的脑机融合系统的研究

发布时间：2018-01-15 15:19

本文关键词：基于迁移学习的脑机融合系统的研究　出处：《浙江大学》2017年博士论文　论文类型：学位论文

【摘要】：磨机融合(Brain Machine Integration ),是指通过脑机接口技术,融合生物智能和机器智能的混合智能系统,被认为是二十一世纪最重要的前沿科技领域之一。近年来,随着脑科学和人工智能的发展,脑机融合可以将生物智能(脑)与机器智能(机),通过脑机接口技术进行有机地融合和深度地协作,进而形成比单一生物智能或者单一机器智能,更加强大的混合智能新形态。同时,随着脑机融合的发展和进步,又可以促进脑科学、人工智能、认知科学与临床医学等领域的理论创新和应用突破,在神经康复与动物机器人领域有着重要的研究意义。作为脑机融合的重要组成部分,机器智能具有强大的存储和运算能力;与机器智能相比,生物智能的优势在于其高效低功耗的感认知和逻辑推理能力。如何将二者的优势有机地融合在一些,建立更强大的新型智能形态,是脑机融合面临的关键问题和挑战。针对这一关键问题和挑战,本文进行了基于迁移学习的脑机融合系统的研究:迁移学习,可以将从不同但相关的领域或者不同但相关的任务中学习到的知识进行迁移和融合;脑机接口,可以在生物大脑与外围设备之间建立直接的连接通路;因此,基于迁移学习和脑机接口的脑机融合系统,可以将不同生物、不同领域、不同任务之间的信息进行交流、知识进行迁移、智能进行融合。概括来说,本文从以下三个方面逐层深入地进行了探讨。首先,本文提出基于迁移学习的脑机融合系统的概念和体系结构,总体思路为:首先,模仿生物的学习过程,使机器具有能够在不同但相似的领域中,解决不同但相关的问题的能力,这对于神经调控、残障康复等缺乏足够多高质量训练数据的领域具有重要的意义;其次,通过脑机融合系统,将学到的知识在生物与生物、生物与机器、机器与机器之间进行迁移和融合,增强系统智能决策的能力,实现大脑-机器-机器-大脑之间深度协作的智能增强系统;此外,通过脑机融合中计算理论与方法的创新,可以为生物大脑运行机制的探索提供新的思路和方法,促进脑科学、认知科学和临床医学的进步。然后,基于本文提出的脑机融合系统的体系结构,针对动物机器人这一重要研究对象,借助浙江大学的大鼠机器人平台,本文设计了基于迁移强化学习的大鼠机器人脑机融合系统。首先,将大鼠机器人迷宫导航问题,抽象为经典的强化学习模型;然后,根据源智能体和目标智能体是否相同、源迷宫和目标迷宫是否相同、源任务和目标任务是否相同,设计了基于层次化的迁移强化学习算法、基于策略复用的迁移强化学习算法、基于值函数复用的迁移强化学习算法和基于规则复用的迁移强化学习算法;接着,基于迁移强化学习算法,从迁移什么、如何迁移、何时迁移三个方面,详细地描述了大鼠机器人脑机融合系统的设计与实现;并从行为实验的角度,证明了基于迁移强化学习的大鼠机器人系统的智能增强性;最后,本文从计算神经建模的角度,解释了此脑机融合系统智能增强的神经机理。最后,本文进一步将基于迁移学习的脑机融合系统的研究,从以动物为对象的实验室研究,拓展到以人类为对象的临床医学的研究。借助哈佛大学的临床诊断和康复平台,本文设计了基于迁移极限学习机的意识诊断和调控脑机融合系统。首先,将大脑意识诊断和调控的问题,抽象为基于皮层脑电的清醒预测和药物控制模型;然后,针对临床医学中高质量数据不足的问题,本文设计了基于特征和参数的迁移极限学习机算法;接着,基于迁移极限学习机算法,本文设计了意识诊断和调控的脑机融合系统;并从临床实验的角度,评估了基于迁移极限学习机的人脑意识诊断和调控系统的有效性;最后,基于此迁移脑机融合系统,本文发现了人脑意识清醒与α震荡具有相关性,并对此神经机理进行了探讨。综上所述,从基于迁移强化学习的大鼠机器人脑机融合系统的研究,到基于迁移极限学习机的人脑意识诊断和调控脑机融合系统的研究,本文逐渐深入地论证了基于迁移学习的脑机融合系统的可行性和有效性;并且,从计算神经建模的角度,解释了基于迁移学习的脑机融合系统智能增强的神经机理;此外,基于设计的脑机融合系统和实验结果,探讨了大脑意识改变的神经机理。本研究为脑机融合系统,在动物机器人和神经康复领域中的发展和应用,提供一种新的思路和方法。
[Abstract]:Mills (Brain Machine Integration), fusion refers to the brain machine interface technology, hybrid intelligent system integration of biological intelligence and machine intelligence, is considered to be one of the most important twenty-first Century in cutting-edge technology. In recent years, with the development of the brain science and artificial intelligence, brain machine fusion can be intelligent (brain) and machine intelligence (machine), organic integration and depth of cooperation through brain computer interface technology, and the formation of biological intelligence than single or single machine intelligence, a new form of hybrid intelligence more powerful. At the same time, along with the development and progress of brain computer integration, but also can promote brain science, artificial intelligence, cognitive science and clinical medicine the application of the theory innovation and breakthrough, has important significance in neural rehabilitation and animal robot field. As an important part of the brain machine integration, intelligent machine has powerful storage And operation ability; compared with machine intelligence, biological intelligence advantage lies in its high efficiency and low power consumption perception and logical reasoning ability. How will the organic integration of the two advantages in some, the establishment of new intelligent form more powerful, is the key problem faced by brain fusion and challenges. Aiming at the key problems and this paper studied the challenge of brain machine transfer learning based on fusion: transfer learning, can be from different but related fields or different but related tasks to learn the knowledge of migration and fusion; brain computer interface, you can establish a connection between the direct pathway of biological brain and peripheral equipment; therefore, system brain machine transfer learning and fusion based on brain computer interface, can be different in different areas, creatures, communicate information between different tasks, knowledge transfer, integration of intelligence. In general, this article from the The following three aspects are discussed deeply. Firstly, this paper puts forward the concept and system structure of brain machine transfer learning based on fusion, the general idea is: first of all, the learning process of imitating biology, so that the machine has a similar but in different areas, different but related problem solving ability, the for neural regulation, plays an important role in disability and lack of enough high quality training data; secondly, through brain computer fusion system, the learned knowledge in biology and biological, biological and machine, migration and integration between machine and machine, enhance the ability of intelligent decision system, intelligent - brain the depth of cooperation enhancement system brain - machine - machine; in addition, the innovation of theory and methods by computing the brain machine fusion, explore the operation mechanism of the biological brain can provide new ideas and methods to promote. Brain science, cognitive science and clinical medicine progress. Then, the system architecture of brain machine is presented in this paper based on the fusion of the animal robot, which is an important research object, with the help of the rat robot platform of Zhejiang University, this paper designs enhance the migration of rat brain machine robot learning system based on fusion. Firstly, the rat robot the maze navigation problem, the abstract for reinforcement learning the classical model; then, according to the source and the target of intelligent agent is the same, the source and target is the same as the maze maze, the source and target tasks are the same, the design of the migration of hierarchical reinforcement learning algorithm based on migration strategy multiplexing reinforcement learning algorithm based on the transfer of value function multiplexing reinforcement learning algorithm and migration rule reuse based on a reinforcement learning algorithm based on migration; then, a reinforcement learning algorithm based on migration from what, how to transfer, When the transfer of three aspects, a detailed description of the design and implementation of the system of rat brain machine robot fusion; and from the behavior experiment angle, demonstrate the enhancement of intelligent robot system enhance the migration of rat based on learning; finally, this paper calculated from neural modeling perspective, explained the neural mechanism of brain machine fusion system intelligent enhancement. Finally, this paper will further study the system of brain machine transfer learning based on fusion, from the animal laboratory of the research object, to expand the clinical medicine research object to humans. By means of clinical diagnosis and rehabilitation platform of Harvard University, this paper designed the diagnosis and control of consciousness brain machine fusion system migration limit based on machine learning. First of all, the problem of consciousness diagnosis and regulation of the brain, as a model to predict and control the abstract awake cortical EEG based on drugs; then, according to the clinical medicine in high quality The problem of lack of data, this paper designed a machine learning algorithm based on the features and parameters of the migration limit; then, the migration algorithm based on extreme learning machine, this paper designs the system of diagnosis and control of consciousness brain machine fusion; and from clinical experimental perspective, the evaluation of the effectiveness of human consciousness diagnosis and control system of the migration of extreme learning machine finally, based on this migration; brain machine fusion system based on this paper, found a correlation between human consciousness and a concussion, and the neural mechanism was discussed. To sum up, from the research system of rat brain machine learning robot fusion based on System Research to strengthen the migration, human consciousness diagnosis and regulation of brain machine migration extreme learning machine the fusion based on this paper has gradually demonstrated the feasibility of the system of brain machine transfer learning and effective fusion based on neural computation and modeling; from the angle of solution The release mechanism of nerve system intelligent enhanced brain machine transfer learning based fusion; in addition, the system and the experimental results of brain machine design based on the combination of neural mechanism of brain consciousness change. This study fusion system for brain machine development and application in animal robot and rehabilitation in the field, to provide ideas and methods new.

【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：R318;TP242

【参考文献】