GraB: Finding Provably Better Data Permutations than Random Reshuffling

Authors: Yucheng Lu, Wentao Guo, Christopher M. De Sa

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

Reproducibility Variable Result LLM Response
Research Type Experimental We show empirically1 on applications including MNIST, CIFAR10, Wiki Text and GLUE that Gra B can outperform random reshuffling in terms of both training and validation performance, and even outperform state-of-the-art greedy ordering while reducing memory usage over 100 .
Researcher Affiliation Academia Yucheng Lu, Wentao Guo, Christopher De Sa Department of Computer Science Cornell University {yl2967, wg247, cmd353}@cornell.edu
Pseudocode Yes Algorithm 1 Herding with Greedy Ordering
Open Source Code Yes The experimental code is available at https://github.com/Eugene LYC/Gra B.
Open Datasets Yes We show empirically1 on applications including MNIST, CIFAR10, Wiki Text and GLUE that Gra B can outperform random reshuffling in terms of both training and validation performance, and even outperform state-of-the-art greedy ordering while reducing memory usage over 100 .
Dataset Splits No The paper mentions using training and validation data, but does not specify explicit dataset split percentages, sample counts, or refer to predefined splits in the main text.
Hardware Specification Yes All the experiments run on an instance configured with a 4-core Intel(R) Xeon(R) 2.50GHz CPU, 32GB memory and an NVIDIA Ge Force RTX 2080 Ti GPU.
Software Dependencies No The paper mentions 'Py Torch' as an example of an ML library, but does not provide specific version numbers for any software dependencies.
Experiment Setup No Detailed information regarding models, datasets and hyperparameters can be found in Appendix A.