毕业生知识库

传统人工神经网络（Artifitial Neural Networks，ANN），特别是以深度学习（Deep Learning，DL）为代表的模式识别方法，需要基于大规模的数据，通过不断迭代训练降低预定义损失函数，以对神经元间的连接权重进行调整，从而完成学习任务。然而，从生物学原理上来说，人类的认知机理并不与此相符，从根本上来说还无法解释人类仅仅需要小样本的输入即可习得新知识的现象。而如何更好地对动作序列进行抽象、获取并表征蕴含于其中的过程性知识也是现今传统人工神经网络并不能很好解决的问题。

本文主要关注类人学习的序列抽象过程，研究类脑智能算法在动作序列分析等相关任务中获取过程性知识的应用，研究的重点在于小样本训练的序列抽象以及基于类脑皮层的模式识别算法。通过调研关于人脑皮质处理信息时的宏观及微观机制，结合受大脑皮层机理启发的层次时序模型（Hierarchical Temporal Memory，HTM）和离散Hopfield网络（Discrete Hopfield Neural Network），并应用于过程性知识表征、动作序列分析及理解等任务中。

本文依据类脑皮层智能算法，实现了抽象模式序列的小样本抗噪学习，提出了一种针对现实场景中机器人的关节运动捕捉数据进行关键帧提取的算法，并且基于对大脑皮层友好的二值化表示，可在原子级别上对划分后的动作序列进行分类。本文从实践中验证了当前类脑皮层模型具有对模式序列进行概念化学习与记忆的能力，提出的关键帧算法能够清晰地提取出运动捕捉轨迹中的关键点，并可以此用于动作场景分析与自然语言描述生成等更高层次的过程性知识获取任务中。

Traditional artifitial neural networks (ANN), especially as deep learning (DL) methods, can get good performance on several pattern recognition tasks with large amounts of data available during training. By iterative training, predefined loss function is gradually reduced, and connection weights between neurons get tuned dynamically, so as to complete the learning task. However, from the perspective of biology, human cognitive mechanism is not consistent with that. Fundamentally, it can not explain why humans can acquire new knowledge only through a few input samples. And also it is still hard for traditional neural networks to extract, characterize and abstract process knowledge contained in action sequences.

This paper mainly focuses on the sequential abstraction process of human-like learning, and studies the application of brain-inspired algorithm in process knowledge acquisition tasks such as action sequence analysis. The research focuses on sequential abstraction of small training set and pattern recognition algorithms based on the cerebral cortex. Through the investigation on the information processing mechanism of cerebral cortex via macro and micro scale, we combined two brain-inspired models Hierarchical Temporal Memory (HTM) and Discrete Hopfield Network, which is applied to process knowledge characterization and action scene understanding.

Based on several brain-inspired algorithms, this paper realizes an anti-noise learning model for small-scale abstract pattern sequences, and puts forward a key frame extraction algorithm for joint motion capture data of a real-scene robot. Furthermore, the action sequences after partition can be classified at atomic level in the form of cortical-friendly binary representation. This paper proves current brain-inspired cortical model with the ability of memorization and conceptual learning of pattern sequences through practice. The key frame algorithm can clearly extract key points in certain motion capture trajectory, and can be applied to more high-level process knowledge acquisition tasks such as action scene analysis and natural language description generation.