Anomaly detection is the process of identifying abnormal events. The proper identification of anomalies can be helpful for almost any domain and is critical in achieving level 4 and 5 autonomy for self-driving cars. However, classical anomaly detection is principally rooted in point-based anomalies, those anomalies that occur at a specific point, which are only a small subset of the more general range-based anomalies that occur in real-world systems.
In this talk, we will discuss necessary advances in anomaly detection to make it relevant in today’s range-based systems. We will also briefly discuss peripheral issues related to anomaly detection, such as neural network verification and the importance of training data.