Mean-field Underdamped Langevin Dynamics and its Spacetime Discretization
Authors: Qiang Fu, Ashia Camage Wilson
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method N-ULA in training a two-layer mean-field neural network to approximate a Gaussian function... Codes of our experiments are available at https://github.com/Qiang Fu09/NULA. More details of the experimental settings and discussion are postponed to Section F. |
| Researcher Affiliation | Academia | 1Department of Computer Science, Yale University, New Haven, CT, USA 2Department of Electrical Engineering and Computer Science, MIT, Cambridge, MA, USA. |
| Pseudocode | Yes | Algorithm 1 (N-ULA) Require: F satisfies Assumptions 2.3-2.7 and 2.9 |
| Open Source Code | Yes | Codes of our experiments are available at https://github.com/Qiang Fu09/NULA. |
| Open Datasets | No | The paper mentions using 'dataset (ai, bi)n i=1' or 'randomly generated data samples' in its examples and experiments, but does not specify a publicly available dataset with a link, DOI, or formal citation. |
| Dataset Splits | No | The paper describes using 'n randomly generated data samples' but does not provide specific details on training, validation, or test splits (e.g., percentages, sample counts, or a description of the splitting methodology). |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running experiments, such as GPU or CPU models, or memory specifications. |
| Software Dependencies | No | The paper does not provide specific version numbers for any software dependencies or libraries used in the experiments. |
| Experiment Setup | Yes | We give the actual updates of the methods involved in our experiment and provide the precise value of parameters in Table 2... We choose K = 104 and also fine-tune h1, λ1 and h2, λ2 to make fair comparison. We postpone our choice of hyperparameters to the Appendix F.1. For each algorithm in our experiment, we initialize xj 0 N(0, 10 2Id) and vj 0 N(0, 10 2Id) for j = 1, ..., N |