Deep learning based anomaly detection in time series and sustainable AI

Room A008

In the ever-evolving landscape of machine learning, detecting concept drift and anomalies is crucial for maintaining model reliability and decision-making accuracy, particularly in dynamic real-world environments. Concept drift, where data distributions change over time, can significantly degrade model performance, while anomaly detection plays a vital role in identifying rare and potentially critical events across various domains.

In this talk, I will present ADDM (Autoregressive-Based Drift Detection Method), a novel approach that effectively identifies and adapts to data shifts, ensuring models remain robust over time. Additionally, I will introduce AnoRand, a deep learning-based semi-supervised anomaly detection method that leverages synthetic label generation to enhance detection performance, even with limited labeled data. Furthermore, I will discuss practical applications in predictive maintenance, focusing on embedded systems.

Finally, I will discuss the growing computational demands of AI models and the resulting CO₂ emissions. As model complexity and training durations increase, so does the energy consumption required for processing.