Fixed-Distance Hamiltonian Monte Carlo
Authors: Hadi Mohasel Afshar, Sally Cripps
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We compare the performance of FDHMC against 3 baselines: (a) static HMC (Neal, 2011), and two versions of No-U-turn sampler (Hoffman and Gelman, 2014) (NUTS) including (b) Dynamic Slice HMC (DSHMC), which is the original implementation (Hoffman and Gelman, 2014), and (c) Dynamic Multinomial HMC (DMHMC) (Betancourt, 2017), which is the current version of NUTS used in STAN (Carpenter et al., 2017). |
| Researcher Affiliation | Collaboration | Hadi Mohasel Afshar CSIRO s Data61 Eveleigh, NSW 2015, Australia Hadi.Afshar@data61.csiro.au Sally Cripps CSIRO s Data61 & The University of Sydney Eveleigh, NSW 2015, Australia Sally.Cripps@data61.csiro.au |
| Pseudocode | Yes | Algorithm 1: FDHMC q(c), U |
| Open Source Code | Yes | Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] |
| Open Datasets | Yes | The posterior density function, πQ(α, β|X, y) (15), associated with Bayesian Logistic Regression for Binary classification where each data point contains a vector of predictors, xi, and the class label yi { 1, 1}: We set σ2 = 1 and apply the model to the following three data sets that are available from the UCI repository (Frank et al., 2011): (a) Australian Credit Approval (Aus Cr) data set where each data point have 14 predictors (as such, n = 15) and we only use the first 100 data points in the data set. (b) SPECT Heart (SPECT) data set with 267 data points, each with 22 predictors. (c) German Credit (Gr Cr) data set with 1000 data points, each with 24 predictors. |
| Dataset Splits | No | The paper does not explicitly provide training/validation/test dataset splits for the datasets used. It describes the length of MCMC chains and burn-in draws: 'The length of each chain is 1000 samples which are drawn after an initial 200 burn-in draws.' |
| Hardware Specification | No | The paper explicitly states in its author checklist that it did not include 'the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [No]'. |
| Software Dependencies | No | The paper mentions software like 'Mici probabilistic programming language (Graham, 2019)' and 'STAN (Carpenter et al., 2017)' but does not provide specific version numbers for these or other software dependencies. |
| Experiment Setup | Yes | For tuning HMC, the initial total simulation duration, λ = ϵL, is set to 2.0 and the step-size is tuned by dual averaging (Nesterov, 2009; Hoffman and Gelman, 2014). ... We tune the the parameters of FDHMC by an algorithms that is presented in the supplementary material and works well on our experimental models. This includes tuning the step-size, ϵ, via dual averaging and the following heuristic for tuning the fixed distance, D: ... The length of each chain is 1000 samples which are drawn after an initial 200 burn-in draws. |