NETS: A Non-equilibrium Transport Sampler
Authors: Michael Samuel Albergo, Eric Vanden-Eijnden
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In experiments on standard benchmarks, highdimensional Gaussian mixtures, and statistical lattice field theory models, NETS shows compelling performances. (...) 3. Numerical Experiments. In what follows, we test the NETS method, for both the PINN objective (25) and the action matching objective (119), on standard challenging sampling benchmarks. |
| Researcher Affiliation | Collaboration | 1Society of Fellows, Harvard University, Cambridge, MA, USA 2NSF AI Institute for Artificial Intelligence and Fundamental Interactions, Cambridge, MA, USA 3Machine Learning Lab, Capital Fund Management, Paris, France 4Courant Institute of Mathematical Sciences, New York, NY, USA. |
| Pseudocode | Yes | Algorithm 1 Training: Note that the resultant set of walkers across time slices {xi k} are detached from the computational graph when taking a gradient step (off-policy learning). 1: Initialize: n walkers, x0 ρ0, A0 = 0, K time steps, model parameters for {ˆbt, ˆFt}, diffusion coefficient ϵt, learning rate η 2: repeat 3: Randomize time grid: t0, t1, . . . , t K Uniform(0, T), sort such that t0 < t1 < < t K 4: for k = 0, . . . , K do 5: tk = tk+1 tk, 6: for each walker i = 1, . . . , n do 7: xi k+1 = xi k + [ˆbtk(xi k) ϵtk Utk(xi k)] tk + 2ϵtk(W i tk+1 W i tk) 8: Ai k+1 = Ai k + [ ˆbtk(xi k) t Utk(xi k) ˆbtk(xi k) Utk(xi k)] tk 9: Estimate (30) by replacing the expectation by an empirical average over the n walkers and the time integral by an empirical average over t0, . . . , t K. 10: Take gradient descent step to update the model parameters. 11: until converged |
| Open Source Code | No | The paper does not provide any explicit statement about open-sourcing their code, nor does it include a link to a code repository. |
| Open Datasets | Yes | A common benchmark for machine learning augmented samplers that originally appeared in the paper introducing Flow Annealed Importance Sampling Bootstrap (FAB) (Midgley et al., 2023) is a 40-mode GMM in 2-dimensions... We next test NETS on Neal s funnel, a challenging synthetic target distribution... as well as the 50dimensional Mixture of Student-T (Mo S) distribution used in (Blessing et al., 2024)... We study the lattice φ4 theory in D = 2 spacetime dimensions (Vierhaus, 2010; Albergo et al., 2019). |
| Dataset Splits | No | The paper describes generating samples from specified target distributions (e.g., 2000 generated samples for ESS/W2) rather than using predefined training/test/validation splits of a static dataset in the traditional sense of supervised learning. |
| Hardware Specification | No | The paper does not specify the hardware (e.g., GPU/CPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions using the 'Python Optimal Transport library' but does not provide a specific version number. No other specific software dependencies with version numbers are listed. |
| Experiment Setup | Yes | We train a simple feed-forward neural network of width 256... We use 100 sampling steps for both, with diffusion coefficients given in the caption of Table 2. ...Diffusion coefficient ϵt = 5, 4 was used for NETS-AM on the Funnel and Mo S, respectively. Equivalently, ϵt = 5, 5 were used by NETS-PINN... We construct 8-mode target GMMs in d = 36, 64, 128, 200 dimensions and learn ˆb with the PINN loss in each scenario. We use the same feed forward neural network of width 512 and depth 4 to parameterize both ˆb and ˆF for all dimensions tested and train for 4000 training iterations. ...The number of sampling steps used to discretize the SDEs in these experiments ranged from K = 100 for ϵt = 0 up to K = 2000 for ϵt = 80. |