Back

Speaker "Alexander Lavin" Details Back

 

Topic

Detecting Anomalies in Streaming Data, Evaluating Algorithms for Real-world Use

Abstract

Much of the world’s data is becoming streaming, time-series data, in domains such as finance, IT, security, medical, energy, and social media. Finding anomalies in time-series data can be critical for many applications, but how do we measure the effectiveness of anomaly detection algorithms for real-world use on streaming data? Traditional benchmarks are batch focused and do not apply to streaming applications. This calls for a scoring framework that does not allow lookahead, rewards early detection, and incorporates continuous learning. Fulfilling this need is the Numenta Anomaly Benchmark (NAB). The goal for NAB is to provide a controlled and repeatable environment of open-source tools to evaluate anomaly detection algorithms on streaming data. NAB includes a custom scoring methodology and a corpus of real-world, labeled data. These components are presented along with results and analyses for several open source, commercially-used algorithms.

Profile

Software and research engineer at Numenta, building machine intelligence by reverse-engineering the neocortex; specializes in anomaly detection and natural language processing (NLP). Lavin studied mechanical engineering at Cornell and Carnegie Mellon Universities, focusing on spacecraft engineering.