Dequantified Diffusion-Schrödinger Bridge for Density Ratio Estimation

Authors: Wei Chen, Shigui Li, Jiacheng Li, Junmei Yang, John Paisley, Delu Zeng

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that D3RE improves robustness and efficiency in downstream tasks such as density ratio estimation, mutual information estimation, and density estimation. Fig. 1 illustrates a comparison of interpolation strategies among DI, DBI, DDBI, and DSBI... We trained DRE(baseline) and D3RE (ours) on eight 2-D synthetic datasets... for 20,000 epochs... The density estimation results are shown in Fig. 2. ...We applied the proposed D3RE framework for density estimation on the MNIST dataset... The results are reported in bits-per-dimension (BPD). Results in Tab. 2 show that D3RE achieves the lowest BPD values... The ablation study on γ2 for density estimation (Fig. 4) reveals systematic trade-offs in performance...
Researcher Affiliation Academia 1The School of Mathematics, South China University of Technology, Guangzhou 510006, China 2The School of Electronic and Information Engineering, South China University of Technology, Guangzhou 510006, China 3The Department of Electrical Engineering, Columbia University, New York, NY 10027, USA. Correspondence to: Delu Zeng <EMAIL>.
Pseudocode Yes Algorithm 1 Training and estimation of D3RE with DDBI Algorithm 2 Training and estimation of D3RE with DSBI
Open Source Code Yes Code is available at https://github.com/Hoemr/Dequantified Diffusion-Bridge-Density-Ratio-Estimation.git.
Open Datasets Yes We trained DRE(baseline) and D3RE (ours) on eight 2-D synthetic datasets, including swissroll, circles, rings, moons, 8gaussians, pinwheel, 2spirals, and checkerboard... We applied the proposed D3RE framework for density estimation on the MNIST dataset...
Dataset Splits No The paper mentions using "2-D Synthetic Datasets" and "MNIST dataset" for experiments. However, it does not explicitly provide specific training/test/validation splits (e.g., percentages, sample counts, or citations to predefined splits) for these datasets in its main text or appendices for its own experiments. For MNIST, it mentions leveraging "pre-trained energy-based models (EBMs) (Choi et al., 2022)" but does not detail the splits used for training their D3RE models.
Hardware Specification No The paper does not explicitly describe any specific hardware used for running its experiments, such as GPU models, CPU models, or cloud computing specifications.
Software Dependencies No The paper mentions using the "Adam optimization method" and the "scalable Sinkhorn algorithm (Cuturi, 2013)". However, it does not provide specific version numbers for any software dependencies, programming languages (e.g., Python), or libraries (e.g., PyTorch, TensorFlow).
Experiment Setup Yes For experiments involving D3RE, we implement both DDBI and DSBI. Unless specified otherwise, we use the following settings: αt = 1 t, βt = t, γ2 = 0.5, ε = 1e 5 and λ(t) = γ2t(1 t). Under this configuration, the interpolant I(X0, X1, t) = (1 t)X0 + t X1 aligns with the Benamou-Brenier solution to the optimal transport problem in Euclidean space (Mc Cann, 1997). The parameterized score model is trained with time score matching loss L3 and optimized with Adam optimization method. ... We trained DRE(baseline) and D3RE (ours) on eight 2-D synthetic datasets... for 20,000 epochs using the joint score matching loss (details in Appendix C.2). ...The batch size is set to 512 for d = {40, 80, 160} and 256 for d = 320, with iteration steps of {40k, 100k, 400k, 500k}, respectively.