Asymptotically Optimal Exact Minibatch Metropolis-Hastings

Authors: Ruqi Zhang, A. Feder Cooper, Christopher M. De Sa

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we show Tuna MH outperforms other exact minibatch MH methods on robust linear regression, truncated Gaussian mixtures, and logistic regression. 5 Experiments We compare Tuna MH to MH, TFMH, SMH (i.e. TFMH with MAP control variates) and Fly MC.
Researcher Affiliation Academia Ruqi Zhang Cornell University rz297@cornell.edu A. Feder Cooper Cornell University afc78@cornell.edu Christopher De Sa Cornell University cdesa@cs.cornell.edu
Pseudocode Yes Algorithm 1 Stateless, Energy-Difference-Based Minibatch Metropolis-Hastings, Algorithm 2 Tuna MH
Open Source Code Yes We released the code at https://github.com/ruqizhang/tunamh.
Open Datasets Yes Lastly we apply Tuna MH to logistic regression on the MNIST image dataset of handwritten number digits.
Dataset Splits No The paper mentions dataset sizes (e.g., N = 10^6 for Gaussian mixture, MNIST) but does not provide specific train/validation/test split percentages or counts.
Hardware Specification No The paper does not provide any specific hardware details such as GPU/CPU models or types of computing resources used for the experiments.
Software Dependencies No We coded each method in Julia; our implementations are at least as fast as, if not faster than, prior implementations.
Experiment Setup Yes We tune the proposal stepsize separately for each method to reach a target acceptance rate, and report averaged results and standard error from the mean over three runs. We set χ to be roughly the largest value that keeps χC2M 2(θ, θ ) < 1 in most steps; we keep χ as high as possible while the average batch size is around its lower bound CM(θ, θ ). We found this strategy works well in practice.