CASIA OpenIR  > 脑网络组研究
基于情境信号的连续多任务学习
曾冠雄
2019-05-31
页数82
学位类型硕士
中文摘要

    人工神经网络(Artificial Neural Networks, ANN)是处理识别和分类任务的强大工具,因为它们可以学习输入和输出之间复杂的映射关系。但是,目前的神经网络所能学习的映射规则通常是单一且固定的。这限制了网络在复杂和动态变化的环境下工作的能力。当神经网络的输出需要根据情境不断变化时,这就对现有的神经网络提出了很大的挑战。在灵长类大脑中,前额叶皮层(Prefrontal Cortex, PFC)会依据情境信号灵活处理当前的输入,执行不同的任务,另外,与当前大多数人工神经网络不同,灵长类能够连续地逐项学习这些情境依赖的处理规则,从而使得其灵活性能够不断地增加。受大脑这些特性的启发,本文提出了一种新的方法,包括正交权重修改(Orthogonal Weights Modification, OWM)算法和情境依赖处理(Context-dependent Processing, CDP)模块。该方法使神经网络能够基于情境连续地学习不同的映射规则。
    本文通过一系列实验证明,OWM算法可以有效保护神经网络在以往任务中获得的知识,从而有效克服学习过程中的“灾难性遗忘”。使用该方法,网络可以在性能几乎不受干扰的前提下,连续地学习多达数千种不同的映射规则,性能优于已经提出的其他连续学习算法。此外,通过使用CDP模块处理情境信息,网络可以连续地学习情境依赖的不同映射规则,实现在相同的输入情况下灵活地根据当前情境完成截然不同的任务。OWM算法和CDP模块的结合使得以后可能创造出紧凑而又高度灵活的人工智能系统,从而逐渐学习现实世界的复杂规律。希望这样可以促进通用智能的产生。最后,本文在卷积神经网络(Convolutional Neural Networks, CNN)和蓄水池计算(Reservoir Computing, RC)网络中验证了OWM算法的有效性和通用性,进一步证明了该算法的扩展性和实用性。

英文摘要

   Artificial neural networks (ANN) are powerful tools for recognition and classification as they learn sophisticated mapping rules between the inputs and the outputs. However, the rules that learned by the majority of current ANN used for pattern recognition are largely fixed and do not vary with different conditions. This limits the network's ability to work in more complex and dynamical situations in which the mapping rules themselves are not fixed but constantly change according to contexts, such as different environments and goals. In addition, unlike most artificial neural networks in use, primates are able to continuously learn these context-dependent processing rules one by one, so that their flexibility can be continuously increased.Inspired by the role of the prefrontal cortex (PFC) in mediating context-dependent processing in the primate brain, here we propose a novel approach, involving a learning algorithm named Orthogonal Weights Modification (OWM) with the addition of a Context-dependent Processing (CDP) module, that enables networks to continually learn different mapping rules in a context-dependent way. Inspired by these characteristics of the brain, we propose a new method, including OWM learning algorithm and CDP module. This method enables the neural network to continuously learn different mapping rules based on context.
   Through a series of experiments, we have proved that the OWM method can effectively protect the knowledge acquired by neural networks in previous tasks, thus effectively overcoming the ``catastrophic forgetting'' in the learning process. We demonstrate that with OWM to protect previously acquired knowledge, the networks could sequentially learn up to thousands of different mapping rules without interference, and the performance is superior to other continuous learning algorithms that have been proposed. In addition, by using the CDP module to modulate the representation of sensory features, a network could sequentially learn different, context-specific mappings for even identical stimuli. Taken together, these approaches allow us to teach a single network numerous context-dependent mapping rules in an online, continual manner. This would enable highly compact systems to gradually learn myriad of regularities of the real world and would promote the generation of generic intelligence. Finally, we validated the effectiveness of the OWM algorithm in Convolutional Neural Networks (CNN) and Reservoir Computing (RC) networks, and further proved the scalability and general applicability of the algorithm.

关键词前额叶皮层 正交权重修改 情境依赖处理 灾难性遗忘 连续学习 多任务学习
语种中文
七大方向——子方向分类类脑模型与计算
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/23772
专题脑网络组研究
推荐引用方式
GB/T 7714
曾冠雄. 基于情境信号的连续多任务学习[D]. 北京. 中国科学院大学,2019.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
基于情境信号的连续多任务学习.pdf(4369KB)学位论文 开放获取CC BY-NC-SA
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[曾冠雄]的文章
百度学术
百度学术中相似的文章
[曾冠雄]的文章
必应学术
必应学术中相似的文章
[曾冠雄]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

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。