Neural Guided Diffusion Bridges

Authors: Gefan Yang, Frank Van Der Meulen, Stefan Sommer

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate the method through numerical experiments ranging from one-dimensional linear to high-dimensional nonlinear cases, offering qualitative and quantitative analyses. Section 5. Experiments.
Researcher Affiliation Academia 1Department of Computer Science, University of Copenhagen, Universitetsparken 1, 2100 København, Denmark 2Department of Mathematics, Vrije Universiteit Amsterdam, De Boelelaan 1111, 1081HV Amsterdam, The Netherlands.
Pseudocode Yes Algorithm 1 Neural guided bridge training
Open Source Code Yes The codebase for reproducing all the experiments conducted in the paper is available in https://github.com/bookdiver/neuralbridge
Open Datasets No The paper's experiments use mathematical models (Linear Processes, Cell Diffusion Model, Fitz Hugh-Nagumo Model, Stochastic Landmark Matching) which generate data through simulation. The models' definitions and parameters are described or referenced in the paper, meaning there isn't a separate, pre-existing external dataset file requiring a specific link or repository for access, beyond the open-sourced code that generates the simulation data.
Dataset Splits No The paper's experiments involve simulating stochastic processes and generating trajectories (e.g., '25,000 independently sampled full trajectories'). It does not use pre-defined external datasets that would require training, validation, or test splits.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments, such as exact GPU/CPU models, processor types, or memory amounts.
Software Dependencies No The paper mentions using JAX and Julia for implementation, but it does not specify version numbers for these software components or any other libraries that would be necessary for reproducibility.
Experiment Setup Yes The map ϑθ is modeled by a fully connected neural network with 3 hidden layers and 20 hidden dimensions for each layer. The model is trained with 25,000 independently sampled full trajectories of X . The batch size was taken to be N = 50 and the time step size δt = 0.002, leading to in total M = 500 time steps. The network was trained using the Adam (Kingma & Ba, 2017) optimizer with learning rate 0.001.