Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

Path Auxiliary Proposal for MCMC in Discrete Space

Authors: Haoran Sun, Hanjun Dai, Wei Xia, Arun Ramamurthy

ICLR 2022 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we show that our path auxiliary algorithms considerably outperform other generic samplers on various discrete models for sampling, inference, and learning. Our method can also be used to train deep EBMs for high dimensional discrete data.
Researcher Affiliation Collaboration Haoran Sun Georgia Institute of Technology EMAIL Hanjun Dai Google Brain EMAIL Wei Xia Amazon EMAIL Arun Ramamurthy Siemens EMAIL
Pseudocode Yes Algorithm 1: Path Auxiliary Sampler (PAS) and the fast version (PAFS)
Open Source Code Yes The code can be found at https://github.com/ha0ransun/Path-Auxiliary-Sampler.git.
Open Datasets Yes We follow Grathwohl et al. (2021) to train a RBM with 500 hidden units on the MNIST dataset using contrastive divergence(Hinton, 2002). ... We train deep EBMs paramterized by Residual Networks(He et al., 2016) on small binary image datasets using PCD(Tieleman & Hinton, 2009) with a replay buffer(Du & Mordatch, 2019). ... Static MNIST, Dynamic MNIST, Omniglot, Caltech Silhouettes
Dataset Splits No The paper mentions tuning 'the expected path length in {3, 5, 7}, and report the result with the best validation likelihood' but does not specify details about the dataset split used for validation (e.g., percentages or sample counts).
Hardware Specification No The paper does not explicitly describe the specific hardware (e.g., GPU/CPU models, memory) used to run the experiments. It mentions that 'the energy function evaluation and gradient calculation usually dominates the computation' but provides no hardware specifications.
Software Dependencies No The paper mentions that 'both of the methods implemented using pytorch' but does not provide specific version numbers for PyTorch or any other software dependencies.
Experiment Setup Yes We use the same setting as Grathwohl et al. (2021), including the batch size, number of iterations, the PCD hyper parameters, etc. ... For our method, we also tune the expected path length in {3, 5, 7}.