Schedule

Outliers are not always signals, nor always noise, but moments where models and reality misalign. These moments reveal not only what the system fails to capture but also what it assumes to be true. Anomalies sit at a conceptual crossroads: they appear as statistical outliers yet often point toward underlying structure or meaning. As models become more adept at detecting deviations from routine patterns, fundamental questions remain. What exactly qualifies as an anomaly? Is it a failure of the system, or of the model's assumptions? Is it noise, novelty, or something else entirely? This talk treats anomalies not as simple exceptions, but as points of tension where the world resists classification, and our abstractions begin to break down. Through examples in human mobility modeling, it explores how AI systems often confuse chaos for meaning, how regularity can be overfit, and how seemingly insignificant deviations can expose critical system-level blind spots. The talk surveys a range of approaches to anomaly detection in spatial and temporal contexts and examines where they succeed and where they misfire. It calls for a broader interpretive vocabulary, both computational and conceptual, for thinking about anomalies in space, time, and society. Rather than asking only how to detect the unexpected, it asks what our models make invisible, which behaviors are marginalized, and how systems can be designed not just to flag the strange, but to learn from it.