| 主页 | 频道首页 | 本站地图 | 论坛留言 | 合作联系 | 本站消息 | |
科技动态 技术发展 文化研究 生物生态 人的研究 生命起源 基因工程 科学普及 科学探索 专题其他

人工类别学习的神经机制

2010-08-05
事件相关电位,类别归纳,认知控制,晚期正成分,扣带前回,前额叶皮层,类别学习
人工类别学习的神经机制

论文中英文摘要
作者姓名:陈安涛
论文题目:人工类别学习的神经机制
作者简介:陈安涛,男,1973年9月出生,2003年9月师从于西南大学李红教授,于2007年1月获博士学位。

人工类别学习的神经机制中文摘要
类别学习是个体获得知识和技能的主要途径,而基于规则的类别学习或人工类别学习则是规则与概念获得的主要形式。自Bruner, Goodnow et al.(1956)以来,人工类别学习在行为水平上得到了大量研究,提出至少四种基本的研究范式,并在此基础上设计了大量的具体实验任务;而近十年来,以脑部损伤病人为研究对象的神经心理学在讨论类别学习的神经机制方面做出了重要贡献,由于某个部位(比如,额叶)受到损伤的病人无法完成特定形式的类别学习任务,因此可以有力地确定任务要求的心理加工与受到损伤的脑结构之间的关系。

  虽然神经心理学在增进我们对人工类别学习的神经机制上做出了重要贡献,但要获得最终的结果仍然遥远,已有研究仍然存在许多有待改进和完善的地方。目前的情况表明,类别学习的研究范式和实验任务有多种,并且相互间差异很大,更重要的是,虽然都声称研究的对象是类别学习,但得到的研究结果却很少一致,由此而提出来的类别学习理论也相互不同。虽然许多认知活动都可认为包含了类别学习,但类别学习必定有反映其本质特征的加工过程,这在已有研究中没有认真加以考虑。

  鉴于类别学习的复杂性,本文将焦点集中于人工类别学习。我们相信,反映人工类别学习本质特征的加工过程是类别归纳,通过类别归纳个体可获得一般性结论,这是所有类别学习任务的共同特征,通过类别归纳获得的一般性结论在得到确认之前仍是假设,检验假设的过程一般是分类,根据假设结论对某个刺激的类别身份做出判断,如果判断错误则可说明假设是错误的,否则假设得到支持。因此,我们认为分类加工在类别学习中扮演着重要的作用,通过分类加工获得的反馈结果直接肯定或否定类别归纳的结果。无论在类别归纳还是分类的加工过程中,认知控制都是重要的,在这类高级认知加工过程中,认知控制使大脑能够觉察信息中出现的变化或异常,从而根据事先的预期或一般性目的而加以修正,修正的过程通常是通过执行对无关的信息的控制、并集中于有关信息而完成的。

  在上述基础上,我们设计了新的类别学习任务,这个任务有如下几个特点:一,刺激材料均是几何图形、颜色和方向,这可有效排除个体的知识经验的差异;二,任务可以容易地分解成相对独立且完整的子任务,比如类别归纳任务、分类任务和认知控制任务等;三,任务在性质上仍然是类别学习任务,即该任务可使被试能够获得规则,并成为其进一步做分类反应的基础;四,各个子任务要求较低,无论是类别归纳、分类还是认知控制均可在1秒之内完成,这为记录子任务完成期间的事件相关电位(ERP)提供了可能。

  采用高密度ERP技术并结合偶极子源分析,我们分别从纵向和横向两个方面研究了类别学习的神经机制。首先,通过对新的类别学习任务的有效分解,将类别学习分解成类别归纳、分类和认知控制这样的三个相对独立又相互联系的子过程,对类别学习所涉及的各个认知加工过程做横向研究;然后,分别对每个子任务做逐步深入的纵向研究,对三个子过程涉及的ERP成分和的时间过程、各个ERP成分对应的心理过程做了研究。研究结果发现:

①,类别归纳在ERP上的效应反映在LPC上,在480 ms左右到达加工的高峰,源分析结果显示,类别归纳加工的颅内源在MTL附近,其中海马的激活水平在类别归纳任务中明显高于特征匹配任务,鉴于海马与新异联结的形成有关,因此海马在任务完成期间受到激活进一步确定任务的性质为类别归纳。

②,无论是在清晰预期条件下还是在模糊预期条件下,分类加工的ERP效应都从N2开始一直持续到P3结束,相应的源分析结果表明分类加工中存在ACC到MTL的连续激活过程,其中ACC激活与N2成分上的差异对应,而MTL激活与P3成分上的差异对应,这说明分类反应中ACC对异常变化(不符合预期)敏感,ACC受到更强激活意味着非靶刺激的出现,而MTL则在需要调用长时记忆的分类任务中起进一步确认判断的作用。

③,对认知控制的研究发现,任务无关信息会首先引发更大波幅的N2成分,与之对应的颅内源确定在ACC上,随后任务无关信息会进一步引起负方向波幅更大的P3成分,与之对应的颅内源确定在PFC上,这一结果为认知控制领域的冲突监测理论提供了直接的支持,也说明在认知控制中存在ACC→PFC的激活转换,其含义是ACC监测到信息当中存在异常情况,随后激活PFC对这些异常情况进行监控。

在整合本文报告的实验结果与已有研究的结果的基础上,我们提出了人工类别学习的神经机制综合模型,这个模型中的关键结构有ACC、PFC和MTL,视觉信息在枕叶皮层和联络皮层区得到初步的感知觉加工,然后暂时存储在PFC的工作记忆单元中,在PFC中会得到初步的知觉整合,反映在P2和N2成分上。N2成分标志着信息加工进入可控制的阶段,在类别归纳、分类和认知控制三个子过程中都是如此,而N2成分的出现与ACC激活有关,ACC激活意味着大脑察觉信息中存在异常,随后ACC将信息传递到PFC中,在类别归纳中PFC将其中保持的特征信息传递到MTL中,其中海马结构执行归纳加工获得刺激间的共同特征;在分类加工中ACC激活后,PFC获得来自ACC的信息可直接判断刺激是否是靶刺激,反映在P3成分上;由分类结果获得的反馈用来确定保持在PFC中的规则是否正确,如果规则是错误的,则中断有关该规则的维持性联系,并通过注意转换来选择新的规则,并再通过分类加工来确认其是否正确,在注意转换过程中ACC和PFC起主要作用,ACC将规则是否是错误的信息传递到PFC,而PFC则执行注意的转换加工。

本研究具有重要的理论意义和现实意义。在理论意义上,加深了对人工类别学习神经机制的理解,进一步完善了人工类别学习的相关理论,同时在研究方法上也有发展,采用新的研究范式,使复杂的类别学习加工得到明确的分解,在此基础上可望针对各个子任务开展深入的纵向水平研究。在现实意义上,类别学习是个体获得规则和概念的主要认知活动,对它的神经机制的研究对于我们理解和分析人类学习行为有重要意义,对于认知功能异常的神经心理诊断也可提供更具针对性的方案。
关键词:事件相关电位;类别归纳;分类;认知控制;晚期正成分;扣带前回;前额叶皮层

The Neural Mechanism of Artificial Category Learning
Chen Antao
ABSTRACT
The category learning is the main approach of acquiring knowledge and skills, and the based-rule category learning or artificial category learning is the main form of acquiring rule and concept. Since Bruner, Goodnow et al. (1965), on the level of behavioral research, there lots of researches have been conducted for the artificial category learning, and on the basis of at least four kinds of paradigms a number of experimental tasks were designed. And since 1990, the neuro-psychology, in which the subjects were brain lesion patient, contributed to understand the neural mechanism of category learning; because specific brain area was damaged the subjects can not complete specific category learning tasks, thus the result can powerfully ascertain the relationship between the mental processing in the task and the specific brain area damaged.

Although the neuro-psychological studies contributed much to our understanding of neural mechanism of category learning, the final conclusion is still far, and there are many room to improve the past relative researches. The current situation is that the kind of paradigm and experimental task is numerous and they are different very much from each other. More important, there is little consist results from these researches, and the theories based on these results are different from each other. Many cognitive activities are included in the category learning, however, there must exist the processing that reflects the nature of category learning, which is omitted in the past researches.

For the complexity of category learning, in this paper we would focus on the artificial category learning. We believe, the cognitive processing reflecting the nature of category learning is the category induction, by which the individual could acquire a general conclusion, which is the common feature in all category learning tasks. However, before ascertained the general conclusion is still a hypothesis, the processing of checking the hypothesis is categorization, in which the category membership of some stimulus is judged by the hypothesis, if the answer is wrong then the hypothesis is wrong, or the hypothesis will be supported. Therefore, we suppose that the categorization plays an important role in the category learning, for the feedback from categorization would directly affirm or denied the hypothesis from category induction. In the category induction and categorization, the cognitive control is important, in such higher level cognitive processing, cognitive control can detect the change or abnormality in the information and correct the way of thinking according to the anticipation: control the irrelative-task information and focus on the relative-task information.

Based on the above analysis, we design new category learning task. There are many features in the new task: first, the stimuli is geometric figures including color and orientation, which excludes effectively the difference of individual knowledge; second, the task can easily be decomposed into relative independent sub-tasks, such as category induction task, categorization task and cognitive control task etc.; third, the nature of task is still category learning, in which the subject must acquire the rule and give correct judgment according to it; forth, the task is relative easy, the three kinds of sub-tasks could be completed in one second which give the possibility for the ERP recorded during the solving these tasks.

Applying high density ERP technique and dipole source analysis, we studied the neural mechanism of category learning. Firstly, by the effective decomposing the category learning task, the category learning was decomposed into three relative and associative sub-tasks: category induction, categorization and cognitive control, the involving cognitive processes in the category learning were studied in cross aspect. And then, each sub-process would be studied in depth respectively, the ERP components, temporal courses and the corresponding cognitive processes involving in the three sub-processes were studied. The main results are showed in the following:

①,The ERP effect of category induction reflects on the LPC component, the corresponding processing attain the peak at 480 ms, the result of dipole source indicates that the encephalic source is about MTL, in which the active level of hippocampus in the category induction task is obvious higher than that in the feature matching task. Considering the relationship between hippocampus and novel connection formation, the activity of hippocampus in sovlving the task further ascertains that the nature of task is category induction.

②,Whether under clear anticipation or ambiguous anticipation, the ERP effect of categorization begins from N2 component and lasts till P3 component, the corresponding encephalic source result shows categorization processing exists from ACC activity to MTL activity, and the activity of ACC corresponds to N2 difference, and the activity of MTL corresponds to P3 difference. These results suggested that ACC is sensitive to the change in the information (violate the anticipation), the activity of ACC begins stronger indicates that the non-target is coming, and MTL plays an important role in the instrumentation of experience in the long-term memory for ascertaining the membership of stimuli.

③,The findings about cognitive control show that the irrelative-task information would elicit higher amplitude of N2 component whose corresponding encephalic source is about ACC, and the irrelative-task information would further elicit higher amplitude of P3 component in the negative orientation whose corresponding encephalic source is about PFC. These results provide direct support for the conflict monitoring theory of cognitive control, and suppose that there exist an activity transfer from ACC to PFC in the cognitive control, the possible processing is that ACC is responsible for the monitoring the change in the information and PFC subsequently is responsible for control the change.

Based on the integration of the results reported in current and past literatures, we put out a integrated neural mechanism mode of artificial category learning, in this mode the key brain structures include ACC, PFC and MTL. The visual information was processed primarily in the occiput and associated cortex, and then maintained in the working memory unit of PFC, and soon integrated in the level of perception in the PFC which reflects on the P2 and N2 components. The occurrence of N2 component indicates that the information processing has been stepped into the controlled phase, which happens in category induction, categorization and cognitive control three sub-processes. The dipole source results indicates that the occurrence of N2 is related with the activity of ACC which suggests that the brain detects the change in the information, and then the activity of ACC triggers PFC. In the category induction, FPC would transfer the feature information maintained into the MTL, the hippocampus in which conducts the abstract of common features during the category induction. In categorization, ACC would trigger the PFC and the later could judge whether the stimulus is the target stimuli directly according the information from ACC, which reflects on P3. The feedback from categorization would be used judge whether the rule maintained in PFC is correct. If the rule is wrong then the loop maintaining the rule would be disrupted and new rule would be selected through attention switch, and judge it’s membership again by categorization. In the attention switch, ACC and PFC play main role, ACC transfers the information whether the rule is right to PFC and PFC conducts the switch of attention.

This research has important theoretic and practical significance. For the theoretic significance, this research improves further theories related with artificial category learning, and new paradigm and method have been put out. Under the new study paradigm, the complex category learning could be deposed into clear sub-processes, and further in-depth studies could be done on each sub-process. For the practical significance, category learning is the main cognitive activity in acquirement of rule and concept, and the study on the neural mechanism of category learning is significant value for our understanding and analysis of human learning behavior, too, based on the results more effective neural-psychological diagnostic project would be put out for the cognitive function abnormality.
Key words: event-related potential, category learning, categorization, cognitive control, late positive component, Anterior cingulate cortex, Prefrontal cortex
提高记忆力和注意力的方法
世界首个人工生命结构诞生
人体的12个极限
研究发现人类大脑不能同时处理超两项任务
人类行为93%可以预测
中西思维方式比较
磁铁让人“不道德” 影响人类大脑是非判断
人脑的发育
决策中的五大心理陷阱
22条方法优化大脑
人类要对人体本身进行深入研究
美科学家发现人类93%行为可以预测
美国式精神疾病开始全球化
从骗子的思路看人生
我们如何做决定
中国科学家对“恐惧记忆形成机制”有新发现
人类理性对秩序的探究
科学家找到大脑智力中心
人类进化,人类起源,人类基因研究,人的研究报告 人的研究报告1 人的研究报告2

本栏目主要介绍人类的故事,包括人类的进化、人类的起源、人类学、人类研究、人类基因发现、人工类别学习的神经机制等。特别关注有关人与文化方面的研究。

『科学频道首页』 『本栏页首』 『关闭窗口』