Learning Optimal Flows for Non-Equilibrium Importance Sampling

Authors: Yu Cao, Eric Vanden-Eijnden

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental On the computational side, we show how to use deep learning to represent the velocity field by a neural network and train it towards the zero variance optimum. These results are illustrated numerically on benchmark examples (with dimension up to 10), where after training the velocity field, the variance of the NEIS estimator is reduced by up to 6 orders of magnitude than that of a vanilla estimator. We also compare the performances of NEIS with those of Neal s annealed importance sampling (AIS).
Researcher Affiliation Academia Yu Cao Courant Institute of Mathematical Sciences, New York University yucaoyc@outlook.com Eric Vanden-Eijnden Courant Institute of Mathematical Sciences, New York University eve2@cims.nyu.edu
Pseudocode No The paper describes algorithms and methods in text and mathematical formulas but does not include any explicit pseudocode blocks or algorithm listings.
Open Source Code Yes The codes are accessible on https://github.com/yucaoyc/NEIS.
Open Datasets No The paper uses synthetic distributions (Gaussian mixtures, funnel distribution) from which samples are drawn. It does not use pre-defined, publicly available datasets with typical train/validation/test splits, but rather generates data on the fly based on the specified distributions.
Dataset Splits No The paper uses synthetic distributions for sampling rather than fixed datasets with explicit train/validation/test splits. Therefore, specific split percentages or counts for validation are not applicable or mentioned.
Hardware Specification Yes All trainings and estimates of Z1 are conducted on a laptop with CPU i7-12700H; we use 15 threads at maximum.
Software Dependencies No The paper mentions using deep learning techniques and neural networks but does not specify particular software libraries (e.g., PyTorch, TensorFlow) or their version numbers.
Experiment Setup Yes We use the finite-time objective Mt ,t+(b) in (7) with t [ 1, 0], t+ = t + 1. ... In practice, we use a time-discretized version of (8) with 2Nt discretization points, and use the standard Runge-Kutta scheme of order 4 (RK4) to integrate the ODE (4) over t [ 1, 1] using uniform time step ( t = 1/Nt). ... we always use an ℓ-layer neural network with width m for all inner layers...The activation function is chosen as the softplus function...We minimize Mt ,t+(b) with respect to the parameters in the neural network using stochastic gradient descent (SGD)...