發(fā)布時間：2018-01-15 15:19

  本文關(guān)鍵詞：基于遷移學習的腦機融合系統(tǒng)的研究　出處：《浙江大學》2017年博士論文　論文類型：學位論文

  更多相關(guān)文章： 腦機融合 遷移學習 腦機接口 遷移強化學習 遷移極限學習機 動物機器人 大鼠機器人 神經(jīng)康復 意識診斷

【摘要】：磨機融合(Brain Machine Integration ),是指通過腦機接口技術(shù),融合生物智能和機器智能的混合智能系統(tǒng),被認為是二十一世紀最重要的前沿科技領域之一。近年來,隨著腦科學和人工智能的發(fā)展,腦機融合可以將生物智能(腦)與機器智能(機),通過腦機接口技術(shù)進行有機地融合和深度地協(xié)作,進而形成比單一生物智能或者單一機器智能,更加強大的混合智能新形態(tài)。同時,隨著腦機融合的發(fā)展和進步,又可以促進腦科學、人工智能、認知科學與臨床醫(yī)學等領域的理論創(chuàng)新和應用突破,在神經(jīng)康復與動物機器人領域有著重要的研究意義。作為腦機融合的重要組成部分,機器智能具有強大的存儲和運算能力;與機器智能相比,生物智能的優(yōu)勢在于其高效低功耗的感認知和邏輯推理能力。如何將二者的優(yōu)勢有機地融合在一些,建立更強大的新型智能形態(tài),是腦機融合面臨的關(guān)鍵問題和挑戰(zhàn)。針對這一關(guān)鍵問題和挑戰(zhàn),本文進行了基于遷移學習的腦機融合系統(tǒng)的研究:遷移學習,可以將從不同但相關(guān)的領域或者不同但相關(guān)的任務中學習到的知識進行遷移和融合;腦機接口,可以在生物大腦與外圍設備之間建立直接的連接通路;因此,基于遷移學習和腦機接口的腦機融合系統(tǒng),可以將不同生物、不同領域、不同任務之間的信息進行交流、知識進行遷移、智能進行融合。概括來說,本文從以下三個方面逐層深入地進行了探討。首先,本文提出基于遷移學習的腦機融合系統(tǒng)的概念和體系結(jié)構(gòu),總體思路為:首先,模仿生物的學習過程,使機器具有能夠在不同但相似的領域中,解決不同但相關(guān)的問題的能力,這對于神經(jīng)調(diào)控、殘障康復等缺乏足夠多高質(zhì)量訓練數(shù)據(jù)的領域具有重要的意義;其次,通過腦機融合系統(tǒng),將學到的知識在生物與生物、生物與機器、機器與機器之間進行遷移和融合,增強系統(tǒng)智能決策的能力,實現(xiàn)大腦-機器-機器-大腦之間深度協(xié)作的智能增強系統(tǒng);此外,通過腦機融合中計算理論與方法的創(chuàng)新,可以為生物大腦運行機制的探索提供新的思路和方法,促進腦科學、認知科學和臨床醫(yī)學的進步。然后,基于本文提出的腦機融合系統(tǒng)的體系結(jié)構(gòu),針對動物機器人這一重要研究對象,借助浙江大學的大鼠機器人平臺,本文設計了基于遷移強化學習的大鼠機器人腦機融合系統(tǒng)。首先,將大鼠機器人迷宮導航問題,抽象為經(jīng)典的強化學習模型;然后,根據(jù)源智能體和目標智能體是否相同、源迷宮和目標迷宮是否相同、源任務和目標任務是否相同,設計了基于層次化的遷移強化學習算法、基于策略復用的遷移強化學習算法、基于值函數(shù)復用的遷移強化學習算法和基于規(guī)則復用的遷移強化學習算法;接著,基于遷移強化學習算法,從遷移什么、如何遷移、何時遷移三個方面,詳細地描述了大鼠機器人腦機融合系統(tǒng)的設計與實現(xiàn);并從行為實驗的角度,證明了基于遷移強化學習的大鼠機器人系統(tǒng)的智能增強性;最后,本文從計算神經(jīng)建模的角度,解釋了此腦機融合系統(tǒng)智能增強的神經(jīng)機理。最后,本文進一步將基于遷移學習的腦機融合系統(tǒng)的研究,從以動物為對象的實驗室研究,拓展到以人類為對象的臨床醫(yī)學的研究。借助哈佛大學的臨床診斷和康復平臺,本文設計了基于遷移極限學習機的意識診斷和調(diào)控腦機融合系統(tǒng)。首先,將大腦意識診斷和調(diào)控的問題,抽象為基于皮層腦電的清醒預測和藥物控制模型;然后,針對臨床醫(yī)學中高質(zhì)量數(shù)據(jù)不足的問題,本文設計了基于特征和參數(shù)的遷移極限學習機算法;接著,基于遷移極限學習機算法,本文設計了意識診斷和調(diào)控的腦機融合系統(tǒng);并從臨床實驗的角度,評估了基于遷移極限學習機的人腦意識診斷和調(diào)控系統(tǒng)的有效性;最后,基于此遷移腦機融合系統(tǒng),本文發(fā)現(xiàn)了人腦意識清醒與α震蕩具有相關(guān)性,并對此神經(jīng)機理進行了探討。綜上所述,從基于遷移強化學習的大鼠機器人腦機融合系統(tǒng)的研究,到基于遷移極限學習機的人腦意識診斷和調(diào)控腦機融合系統(tǒng)的研究,本文逐漸深入地論證了基于遷移學習的腦機融合系統(tǒng)的可行性和有效性;并且,從計算神經(jīng)建模的角度,解釋了基于遷移學習的腦機融合系統(tǒng)智能增強的神經(jīng)機理;此外,基于設計的腦機融合系統(tǒng)和實驗結(jié)果,探討了大腦意識改變的神經(jīng)機理。本研究為腦機融合系統(tǒng),在動物機器人和神經(jīng)康復領域中的發(fā)展和應用,提供一種新的思路和方法。
[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.

【學位授予單位】：浙江大學
【學位級別】：博士
【學位授予年份】：2017
【分類號】：R318;TP242

【參考文獻】

相關(guān)期刊論文 前10條

1 鄭筱祥;王怡雯;張韶岷;張巧生;;猴子PMd區(qū)腦電解碼抓握手勢及機械手實時控制[J];科技創(chuàng)新導報;2016年12期

2 吳曙霞;蔣麗勇;劉偉;武士華;;DARPA生物技術(shù)研究進展與啟示[J];軍事醫(yī)學;2016年06期

3 李遠清;;腦機接口技術(shù)在意識障礙領域應用的前景展望[J];中華神經(jīng)創(chuàng)傷外科電子雜志;2015年02期

4 吳朝暉;潘綱;;腦科學的新手段新技術(shù):信息+系統(tǒng)+智能視角[J];科學通報;2015年10期

5 莊福振;羅平;何清;史忠植;;遷移學習研究進展[J];軟件學報;2015年01期

6 吳朝暉;俞一鵬;潘綱;王躍明;;腦機融合系統(tǒng)綜述[J];生命科學;2014年06期

7 許敏鵬;張力新;明東;綦宏志;陳龍;馬嵐;萬柏坤;;基于SSVEP阻斷與P300特征的混合范式腦-機接口[J];電子學報;2013年11期

8 王行愚;金晶;張宇;王蓓;;腦控:基于腦-機接口的人機融合控制[J];自動化學報;2013年03期

9 林涵;石海明;曾華鋒;;從DARPA資助BCI技術(shù)研發(fā)看未來軍事變革[J];國防科技;2011年05期

10 王皓;高陽;陳興國;;強化學習中的遷移:方法和進展[J];電子學報;2008年S1期

相關(guān)博士學位論文 前3條

1 倪彤光;基于遷移學習的特征選擇與分類方法及其應用研究[D];江南大學;2015年

2 龍明盛;遷移學習問題與方法研究[D];清華大學;2014年

3 張倩;基于知識表達的遷移學習研究[D];中國礦業(yè)大學;2013年

相關(guān)碩士學位論文 前2條

1 許至杰;遷移學習理論與算法研究[D];華東師范大學;2012年

2 戴文淵;基于實例和特征的遷移學習算法研究[D];上海交通大學;2009年

，

本文編號：1428916