Robust and Versatile Event Detection through Gradient-Based Scoring of HMM Models.

Abstract

Event detection is a critical feature across any data-driven system as it assists with the identification of nominal and anomalous behavior. Event detection is all the more relevant in robotics as robots are required to operate with greater autonomy and in increasingly unstructured environments. Recently supervised and unsupervised machine learning methods along with probabilistic models have played increasingly important roles in modeling and predicting events. In this work, we present an accurate, robust, fast, and versatile measure for skill and anomaly identification. Our work presents a theoretical proof that establishes the link between the derivative of the HMM log-likelihood and the latest emission probabilities. The result used a set of verifiable suppositions and insights aided by the log-sum-exp trick. This link established that the latest emission probabilities directly affected the log-likelihood gradient. The key insight was the inverse relationship: that from these gradients skills and anomalies could be identified. Our results showed better performance across all metrics than related state-of-the art results. The result is very significant for event detection problems as the gradient-based measure can be broadly utilized in any subject area (finance, bioinformatics, natural language processing, computer vision, etc) that involves HMM modeling, including its extensions: parametric, nonparametric, hierarchical, etc).