Deep Learning for Scalable Time Series Classification

This thesis develops scalable and label-efficient deep learning methods for real-world time series data, such as brain activity recordings (EEG) or sensor data. While most existing tools are designed for small, fully labeled datasets, practical applications often involve much larger, noisier data with limited labels. To address this, we introduce new datasets, models, and training strategies that improve accuracy and generalization at scale. Our models are also tailored to handle challenges in EEG data, such as noise and variability across recording devices, enabling more reliable analysis across diverse real-world scenarios.