Abstract

Autonomous streaming anomaly detection can have a significant impact in any domain where continuous, real-time data is common. Often in these domains, datasets are too large or complex to hand label. Algorithms that require expensive global training procedures and large training datasets impose strict demands on data and are accordingly not fit to scale to real-time applications that are noisy and dynamic. Unsupervised algorithms that learn continuously like humans therefore boast increased applicability to these real-world scenarios.

Hierarchical Temporal Memory (HTM) is a biologically constrained theory of machine intelligence inspired by the structure, activity, organization and interaction of pyramidal neurons in the neocortex of the primate brain. At the core of HTM are spatio-temporal learning algorithms that store, learn, recall and predict temporal sequences in an unsupervised and continuous fashion to meet the demands of real-time tasks. Unlike traditional machine learning and deep learning encompassed by the act of complex functional approximation, HTM with the surrounding proposed framework does not require any offline training procedures, any massive stores of training data, any data labels, it does not catastrophically forget previously learned information and it need only make one pass through the temporal data.

Proposed in this thesis is an algorithmic framework built upon HTM for intelligent streaming anomaly detection. Unseen in earlier streaming anomaly detection work, the proposed framework uses high-order prior belief predictions in time in the effort to increase the fault tolerance and complex temporal anomaly detection capabilities of the underlying time-series model. Experimental results suggest that the framework when built upon HTM redefines state-of-the-art performance in a popular streaming anomaly benchmark. Comparative results with and without the framework on several third-party datasets collected from real-world scenarios also show a clear performance benefit. In principle, the proposed framework can be applied to any time-series modeling algorithm capable of producing high-order predictions.

Publication Date

2018

Document Type

Thesis

Student Type

Graduate

Advisor

Dhireesha Kudithipudi

Advisor/Committee Member

Zachary Butler

Advisor/Committee Member

Nathan Cahill

Comments

This thesis is embargoed. The full-text will be available on or around 5-25-2019.

Campus

RIT – Main Campus

Available for download on Saturday, May 25, 2019

Share

COinS