TOTEM - Tokenized Time Series Embeddings for General Time Series Analysis

Watch a 29-minute research presentation exploring TOTEM (TOkenized Time Series EMbeddings), a novel approach to unified time series analysis. Learn about a simple tokenizer architecture that creates discrete, learnt data representations enabling cross-domain training and generalist applications. Discover how TOTEM embeds time series data from various domains using self-supervised learning, demonstrating effectiveness across multiple tasks with minimal tuning. Follow along as Sabera Talukder presents comprehensive evaluation results from 17 real-world datasets across 3 tasks, showing how TOTEM matches or exceeds performance of existing methods in both specialist and generalist settings. Gain insights into this innovative contribution to general time series analysis, with implementation details available through the provided GitHub repository.