TL;DR
This paper introduces a simulation-free neural framework for modeling diffusion processes that directly incorporates the Fokker-Planck equation, enabling efficient training across various applications without costly simulations.
Contribution
It proposes a coupled neural parameterization that enforces density and dynamics constraints, simplifying the construction of Neural Conservation Laws for diverse diffusion-based tasks.
Findings
Effective in modeling spatio-temporal events
Learns optimal dynamics from population data
Works across generative modeling and optimal control
Abstract
We present a novel simulation-free framework for training continuous-time diffusion processes over very general objective functions. Existing methods typically involve either prescribing the optimal diffusion process -- which only works for heavily restricted problem formulations -- or require expensive simulation to numerically obtain the time-dependent densities and sample from the diffusion process. In contrast, we propose a coupled parameterization which jointly models a time-dependent density function, or probability path, and the dynamics of a diffusion process that generates this probability path. To accomplish this, our approach directly bakes in the Fokker-Planck equation and density function requirements as hard constraints, by extending and greatly simplifying the construction of Neural Conservation Laws. This enables simulation-free training for a large variety of problem…
Peer Reviews
Decision·UAI 2025 Poster
Personally, I like the text. It is more-or-less clearly written, understandable (but somewhere in pages 5-6 some inconsistencies appear with indices ($i \leftrightarrow D$) - I reflected them in my questions / comments); logic is clear. So, I thank the authors (but some of the “raw” moments need tweaking). Also, It was a pleasant nostalgia for me, when reading about all of these autoregressive models and mixtures of logistics.
1. The main question for the method (and it is partially reflected in the appendix) is that the approach is not dimension-scalable. Also, the proposed parameterization with autoregressive models and mixture of logistics does not seem to be very expressive. 2. No source code for the submission. 3. While text is clearly written, some "raw" moments and misprints to be fixed (se below). Also, please see my questions and comments below
* The idea of converting the constrained problem into an unconstrained one is nice. * The method is also theoretically justified, although I have not fully checked the math. * The proposed approach is "simulation-free", which I think should translate to better cost-accuracy tradeoff.
### Major * The method claims scalability (i.e. better cost-accuracy trade-off) as a primary benefit, yet there is no such comparison (experimental or otherwise). A comparison of the wall-clock time or FLOPS will be good. * While this is the first simulation-free approach to address Section 4.3, is there a non-simulation-free baseline to check the quality of the solution? * Justifications for Lemma 3 and Theorem 1 are missing. Remarks on Proposition 1 would be helpful, although this might be tri
- Hard-coding PDEs or PDE terms often outperforms soft penalties (as would be prevalent in physics-informed neural networks). Constructing an ansatz that does this for a Fokker-Planck equation is thus quite interesting. - The paper is well-written and more or less understandable, even for me as a reader outside of the field.
- I am unfortunately not very familiar with the challenges or difficulties of the field, so I will pose questions rather than pointing out weaknesses.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
