Recent Advances in Time Series Foundation Modeling

Time series foundation models are changing how we work with time-based data, similar to how large pre-trained language and vision models transformed text and image modeling. In this talk, we will briefly introduce MOMENT, one of the first families of open-source time series foundation models. MOMENT models offer three key advantages: they work as building blocks for various time series tasks (like forecasting, classification, anomaly detection, and imputation); they perform well immediately without needing many examples (enabling zero-shot forecasting and few-shot classification); and they can be fine-tuned with in-distribution, task-specific data to improve results. We will show how MOMENT is already being used in real applications. The talk will conclude with our latest research on developing foundation models that can handle longer and multivariate time series, methods to understand and steer these models, and how large language model agents can enhance time series machine learning engineering.

Bios:
Artur Dubrawski, Ph.D. M.Eng, is an Alumni Research Professor Chair of Computer Science at Carnegie Mellon University where he directs the Auton Lab, one of the largest applied machine learning and artificial intelligence teams in academia. For more than 3 decades he has been working on the forefront of development of AI serving in technical leadership roles in industry and leading multiple research endeavors in academia.
Mononito Goswami recently graduated with a Ph.D. in Robotics from Carnegie Mellon University. He is interested in developing foundational machine learning (ML) techniques for real-world applications. His research tackles the limitations of traditional ML approaches, focusing on scenarios with inaccurate, decentralized, and insufficient data, all in effort to democratize ML. He led the development of one of the first open-source foundation models for time series data.