Detecting Anomalies in Streaming Data
summary
An open-source HTM implementation processes each timestamped value in ~10 ms, continuously learns temporal sequences and emits a normalized anomaly score plus a second-order anomaly likelihood that rewards earlier detections; it flags subtle pre-failure patterns that simple thresholds miss and outperforms Twitter and Skyline baselines on a held-out collection. The team releases the NAB benchmark: 58 streams, 365,000 points (47 real labeled streams, 11 synthetic) and a scoring rule that penalizes late detections and false alarms.
AI-written summary · about
chapters (8)
- 0:00Introduction and overview
- 3:41Why anomaly detection matters
- 7:12Real-time streaming constraints
- 11:29HTM for anomaly detection
- 17:30Benchmarking streaming anomaly detection
- 26:39NAB scoring and results
- 33:04Summary and resources
- 42:49Q&A and closing
related talks
39mMastering real-time anomaly detection with open source tools
19mSubutai Ahmad, VP of Research, Numenta
30mRealtime Time Series Anomaly Detection in Production
28mUnsupervised Real-Time Stream-Based Novelty Detection Technique
45mAnomaly Detection in Time Series
43mBeyond Logs: Real-Time Anomaly Detection with Machine Learning
44mAnomaly Detection in Telecom with Spark
41mApplying Foundational Models for Time Series Anomaly Detection