基于层次时态记忆架构的时序模式匹配技术研究

Because of its importance and universality, time series pattern matching has been attracting research interest for many years. However, its development is still not satisfactory. Hierarchical Temporal Memory (HTM) is an innovative computational intelligence model simulating the functionalities of human brain neocortex. It has the advantage of being flexible, stable, and general-purpose. In this project we innovatively study the time series pattern matching problem based on the HTM architecture. We can exploit the huge number of connections between the large number of neurons in the HTM to record the invariant representation of time series which truthfully reflects the similarity between time series pattern under a variety of distortions and noises. We can also implement flexible matching methods based on HTM's ability of storing pattern sequence and predicting future patterns. After optimizing and improving the HTM architecture, we strive to achieve intelligent, flexible, and efficient time series pattern matching technique. The implementation of this project will not only greatly improve the time series pattern matching technique, but also deepen the understanding and mastery of domestic researchers to HTM architecture, and even further develop and improve this architecture.

由于其重要性和普遍性，时序模式匹配研究受到了多年的关注，然而其发展水平还不尽如人意。层次时态记忆（HTM）架构是一个新颖的模拟人类新大脑皮层功能的智能计算模型，具有灵活、稳定、通用的特点。本项目创新性地提出将HTM架构应用于时序模式匹配问题，利用HTM架构中大量神经元之间海量且灵活的连接记录时序模式在各种形变及噪声下的恒定表征，如实地反应出时序模式间的相似性；利用HTM架构对模式序列的记忆及预测能力实现灵活多变的匹配方式；并通过对HTM架构的优化及改进力求实现智能、灵活、高效的时序模式匹配技术。本项目的实施不仅将大大提高时序模式匹配技术的研究水平，还将加深国内研究人员对HTM架构的理解与掌握，乃至进一步的发展和完善这个架构。

由于其重要性和普遍性，时序模式匹配研究受到了多年的关注，然而其发展水平还不尽如人意。层次时态记忆（HTM）架构是一个新颖的模拟人类新大脑皮层功能的智能计算模型，具有灵活、稳定、通用的特点。本课题将HTM架构应用于时序模式匹配，对多项关键技术展开了深入研究，研究成果主要包括：（1）提出将离散数据和实数数据二进制化的方法；（2）提出构造包含多种形变、噪声及反应数据统计特性的训练模式流的方法以训练HTM区域在形变和噪声的干扰下有效识别相似的时序模式；（3）在修改原有HTM算法的基础上提出在HTM层次中有效存储时序数据的方法，这一点也是本研究区别于其他HTM应用的重要之处；（4）提出了支持各种灵活的时序模式匹配方式的算法。但由于HTM的不成熟，上述成果还有一些不完善之处，有待今后进一步的研究和提高。