Tensor Train Diffusion solves high-dimensional sampling 10x faster
Low-rank tensor formats beat PINNs for score-based sampling—no hyperparameter sweat.
Deep Dive
A new solver for score-based diffusion sampling uses functional tensor train (FTT) to exploit low-rank structures in high-dimensional functions, enabling fast, robust, and accurate generation from complex probability densities by solving the underlying Hamilton-Jacobi-Bellman PDE via backward stochastic differential equations.
Key Points
- Replaces expensive PINNs with functional tensor train (FTT) format to solve HJB equation, reducing training time by up to 10x.
- Exploits low-rank structures in high-dimensional functions for model compression and rapid computation.
- Integrates FTT with backward SDEs (BSDEs) for stable, hyperparameter-robust score-based sampling.
Why It Matters
Fast, accurate high-dimensional sampling unlocks new applications in scientific computing, from drug discovery to financial risk modeling.