On the Connection between Pre-training Data Diversity and Fine-tuning Robustness

Authors: Vivek Ramanujan, Thao Nguyen, Sewoong Oh, Ali Farhadi, Ludwig Schmidt

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate our findings on pre-training distributions drawn from various natural and synthetic data sources, primarily using the i Wild Cam-WILDS distribution shift as a test for robustness.
Researcher Affiliation Collaboration Vivek Ramanujan Thao Nguyen Sewoong Oh Ludwig Schmidt Ali Farhadi University of Washington Allen Institute for AI {thaottn,ramanv}@cs.washington.edu
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any explicit statements about releasing source code or links to a code repository for their methodology.
Open Datasets Yes We use Image Net [11] and i Naturalist [46] as the primary pre-training distributions of interest... The downstream task of interest is wildlife classification with the i Wild Cam WILDS dataset [22].
Dataset Splits No There are two test sets for evaluation: ID test data consists of images taken by the same camera traps as the training set, but on different days from the training and validation (ID) images. In contrast, OOD test data contains images taken by a disjoint set of camera traps from training and validation (ID) images.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments, such as GPU or CPU models.
Software Dependencies No The paper does not provide specific version numbers for software dependencies or libraries used in the experiments.
Experiment Setup No The paper mentions 'fine-tuning for 12 epochs' and refers to 'Appendix A [for] further training details', but the main text does not provide a comprehensive set of specific hyperparameters or system-level training settings needed for reproduction.