Seminar LAMBDA: Representation learning methods for optimal change point detection procedures

Abstract:

In sequential data analysis, a change point is a moment of abrupt regime switch in data streams. Detecting these changes quickly and accurately is crucial in various scenarios, from industrial sensor monitoring to challenging video surveillance. However, classic approaches for change point detection (CPD) often struggle with high-dimensional and semi-structured data due to their intricate nature, as they cannot process such data without a proper representation. Moreover, they often work only under strong assumptions and fail to adapt to specific domains with peculiarities. 

In this thesis, we propose a comprehensive framework for CPD via representation learning suitable for a diverse range of scenarios. For low-dimensional industrial data, an introduced two-stage system for effective CPD can work given expert annotations. Alternatively, we have a model that provides universal representations for unsupervised settings. To handle complex high-dimensional data, we propose a novel theoretical-grounded loss function that allows representation learning for deep models. Our loss function balances the trade-off between detection delay and time to false alarm, providing an asymptotic lower bound for established rigorous solutions while remaining differentiable. Under realistic assumptions, it serves as a tight approximation with a power-law of convergence rate. Through comprehensive experiments on synthetic sequences, real-world sensor data, and high-dimensional video datasets, we demonstrate the importance of meaningful representations tailored to the CPD task.

Time: 14:40-16:00 (UTC+3)

Place: Pokrovsky blvr, 11, R505

Lang: Eng