Date: July 29th, 2025

Time: 9:00 AM EST

Location: Coda C1315

Zoom Link: https://gatech.zoom.us/j/99980991336?pwd=AFmePvjXJbsn8TCMg41ivtF9QV2tax.1&from=addon

Harshavardhan Kamarthi

Machine Learning PhD Student

School of Computational Science and Engineering 

Georgia Institute of Technology

Committee

1. Dr. B Aditya Prakash (Advisor)

2. Dr. Chao Zhang

3. Dr. Yao Xie

4. Dr. Thomas Ploetz

5. Dr. Albert Gu

Abstract:

Deep learning has significantly improved time‑series analytics, but existing models often fail to remain reliable and generalizable across domains, offering poor calibration and struggling to scale to high‑dimensional data.  This thesis introduces a suite of frameworks that make large time‑series models more generalizable, reliable, and scalable.  Practitioners will be able to deploy these models in real‑time settings, obtain robust calibrated forecasts despite distribution shifts, and reuse the same data‑efficient models across diverse domains, even in zero or few‑shot scenarios.  The thesis first develops non‑parametric probabilistic methods that capture and propagate uncertainty across multiple time-series and modalities to deliver dependable prediction intervals.  

Next, we construct foundational time‑series models pre‑trained on heterogeneous datasets using novel supervised pre‑training and semantic tokenization strategies, providing reliable performance across domains, and minimizing the need for task‑specific fine‑tuning.  Building on these advances, this thesis scales solutions to applications involving thousands of interrelated series, supporting hierarchical and multivariate forecasting while retaining performance and reliability.  Extensive benchmark tests and industrial deployments validate our approach, pushing the field toward more trustworthy, scalable, and broadly applicable time‑series forecasting systems.