A Reparametrization-Invariant Sharpness Measure Based on Information Geometry

Authors: Cheongjae Jang, Sungyoon Lee, Frank Park, Yung-Kyun Noh

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We now apply our sharpness measure as a regularizer to train neural networks to see if regularizing the measure can improve the generalization performance. The corresponding loss function is written as... We consider a toy example and then consider image classification tasks involving MNIST and CIFAR-10/100 data sets. ... Table 1: Averages and standard errors of the test classification accuracies for SGD, GR, SAM, ASAM, and SGD with our regularization method on MNIST, CIFAR-10, and CIFAR-100 data sets.
Researcher Affiliation Collaboration Cheongjae Jang Hanyang University cjjang@hanyang.ac.kr Sungyoon Lee Korea Institute for Advanced Study sungyoonlee@kias.re.kr Frank C. Park Seoul National University Saige Research fcp@snu.ac.kr Yung-Kyun Noh Hanyang University Korea Institute for Advanced Study nohyung@hanyang.ac.kr
Pseudocode No The paper describes methods and computations but does not include a clearly labeled "Pseudocode" or "Algorithm" block.
Open Source Code Yes 3. If you ran experiments... (a) 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] See supplemental materials.
Open Datasets Yes We consider a toy example and then consider image classification tasks involving MNIST and CIFAR-10/100 data sets.
Dataset Splits No No explicit percentages or sample counts for training/validation/test splits are provided in the main paper. It mentions 'Train Test Poison' in figures, but no specific split details.
Hardware Specification No The paper states 'See Appendix G.4.2.' for hardware specifications, but Appendix G.4.2 is not provided in the given text.
Software Dependencies No All the experiments have been performed using the Py Torch library [38]. (No version specified)
Experiment Setup No For this experiment, we regularize the mini-batch IGS defined in (7)... The detailed experimental settings are explained in Appendix G.4.2.