Proper Losses for Discrete Generative Models

Authors: Dhamma Kimpara, Rafael Frongillo, Bo Waggoner

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we experimentally evaluate our losses as a proof of concept.
Researcher Affiliation Academia 1University of Colorado Boulder.
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks.
Open Source Code No The paper does not provide an explicit statement about releasing source code or a link to a code repository for the methodology it describes.
Open Datasets No The paper conducts numerical experiments using theoretical distributions (e.g., 'power law distributions', 'Zipfians') rather than external public datasets that would require access information. Thus, there is no public dataset access information provided.
Dataset Splits No The paper does not provide specific training, validation, or test dataset splits. The experiments involve drawing samples from distributions for evaluation, not partitioning a fixed dataset for model training and evaluation in the typical machine learning sense.
Hardware Specification No The paper does not explicitly describe the hardware used to run its experiments, such as specific GPU or CPU models.
Software Dependencies No The paper does not provide a reproducible description of ancillary software with specific version numbers (e.g., programming languages or libraries with their versions).
Experiment Setup Yes For each pair of distributions p and q, at each number of total samples, we measured the absolute deviation between the loss value and the true distance between the distributions. We drew up to K1.5 total samples. We repeated this experiment for various batch sizes, where at each iteration, we drew the same batch size from p and q.