Hamiltonian Dynamics with Non-Newtonian Momentum for Rapid Sampling
Authors: Greg Ver Steeg, Aram Galstyan
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We introduce some standard synthetic benchmarks along with variations to help contrast methods. ... To measure how well the samples resemble true samples from the distribution we use the (unbiased) kernel estimator of the squared Maximum Mean Discrepancy (MMD) [43]... We perform our experiments without burn-in in the transient regime, ESS can be interpreted as a proxy for mixing speed, as it depends on the auto-correlation in a chain. ... Results are summarized in Fig. 4 and Table 3.2, with a visualization in Fig. 14. |
| Researcher Affiliation | Academia | Greg Ver Steeg Aram Galstyan University of Southern California, Information Sciences Institute {gregv,galstyan}@isi.edu |
| Pseudocode | Yes | For efficient ergodic sampling with the ESH dynamics, we do not store the entire trajectory but instead use reservoir sampling (Alg. 3). |
| Open Source Code | No | The paper does not contain an explicit statement about open-sourcing their code or a link to a code repository for their methodology. |
| Open Datasets | Yes | For our experiment, we used a setting resulting from the extensive hyper-parameter search in [21]. We train a convolutional neural network energy model on CIFAR-10 using a persistent buffer of 10,000 images. |
| Dataset Splits | No | The paper mentions using CIFAR-10 and various synthetic datasets but does not provide specific details on training, validation, or test dataset splits (e.g., percentages, sample counts, or citations to predefined splits) in the main text for reproduction. |
| Hardware Specification | No | The paper does not specify the exact hardware used for experiments, such as specific GPU or CPU models, memory details, or cloud instance types. |
| Software Dependencies | No | The paper does not explicitly list any software dependencies with specific version numbers (e.g., programming languages, libraries, or frameworks). |
| Experiment Setup | Yes | In experiments we set these to be as large as possible on a logarithmic grid without leading to numerical errors. This leads us to use a value of = 0.1 in subsequent experiments. For the Runge-Kutta integrators we were forced to use rather small sizes for the relative tolerance (10 5) and absolute tolerance (10 6) to avoid numerical errors. |