Sriram Gopalakrishnan, Ph.D.

Sriram Gopalakrishnan, Ph.D.

AI Research Lead at J.P. Morgan Chase | Sequential-Decision Making Problems, Human-AI Interaction, & Generative AI

Constrained Time-Series Diffusion & Multi-Modal Retrieval

Published:

Pillar: Sequential Decision Making
Diffusion Models Time-Series Non-Convex Optimization Multi-Modal Retrieval

Overview

Financial and industrial systems require generating realistic time-series data while strictly obeying hard physics, accounting, and regulation constraints.

Our research yielded two breakthroughs:

  1. Constrained Time-Series Generation (NeurIPS 2023): Blended score-based diffusion models with projected gradient methods to enforce arbitrary non-convex constraint sets during reverse diffusion.
  2. Multi-Modal Time-Series Retrieval (ICAIF 2023 Best Industry Paper Award & US Patent 12,475,133): Created a shared latent space projecting unstructured text/news and continuous multivariate financial time series to enable cross-modal semantic search and anomaly detection.