Semi-supervised anomaly detection
Effective anomaly detection is a tough problem - just because a data point is relatively rare does not mean it is significant enough for further investigation or action in the context of the use case (e.g. threat detection, IT system health, health prediction, etc.). This results in a dilemma - basic unsupervised clustering ML is easy to set up but prone to false positives; alternately, more accurate supervised models struggle with acquiring and updating sufficient labeled training data. Orca’s solution provides the best of both approaches, combining the simple start-up/operation of unsupervised systems with the online-learning and continuously improving accuracy of more supervised models.
Solving the "hard to get labeled data" problem
With Orca you can start with unsupervised anomaly detection - no labels required - and when anomalous events get flagged by a human operator or agent system as significant, they become labeled parts of the model memory. Once a few flagged examples are in the model memory, the system will reveal both how anomalous a given datapoint is and how likely it is to be significant.
Continuous accuracy improvement
Orca models learn in real-time from examples as they are flagged - no feature engineering or rule writing required. This means mistakes (e.g. false positives) can be corrected extremely quickly and the model improves continuously. This online learning improves both the “pure” anomaly detection capabilities and the significance classification ratings. Orca’s hybrid anomaly detection/classification setup, combined with its online learning capabilities means fewer false positives and no zero repeat false-positives.
Orca enables you to create semi-supervised hybrid anomaly-detection/ classification models that give you the best of both supervised and unsupervised approaches while still being very light-weight and able to handle large volumes of data cost effectively.
Find out if Orca is right for you
Speak to our engineering team to learn how we can help you unlock high performance agentic AI / LLM evaluation, real-time adaptive ML, and accelerated AI operations.