Learning to Select Pivotal Samples for Meta Re-weighting

Authors: Yinjun Wu, Adam Stein, Jacob Gardner, Mayur Naik

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical studies demonstrate the performance advantage of our methods over various baseline methods. We further explore whether our methods select reasonable meta samples for re-weighting noisily labeled data and class-imbalanced data by conducting experiments on reweighting MNIST, CIFAR, and Imagenet-10 datasets in the presence of noisy labels or imbalanced class distribution.
Researcher Affiliation Academia University of Pennsylvania wuyinjun@seas.upenn.edu, steinad@seas.upenn.edu, jacobrg@seas.upenn.edu, mhnaik@cis.upenn.edu
Pseudocode Yes In the end, we visually present both RBC and GBC equipped with this sampling technique in Figure 2 and include their pseudo-code in Algorithm 3 in Appendix Details of the adapted K-means algorithm .
Open Source Code Yes All the code is publicly available5. with footnote 5https://github.com/thuwuyinjun/meta sample selections
Open Datasets Yes We demonstrate the effectiveness of our methods for training deep neural nets on image classification datasets, MNIST (Deng 2012), CIFAR-10 (Krizhevsky, Hinton et al. 2009) and CIFAR-100 (Krizhevsky, Hinton et al. 2009), and Imagenet-10 (Russakovsky et al. 2015)4.
Dataset Splits Yes This toy dataset is then divided into 600 training, 240 testing, and 160 validation samples using a random partition. We report the validation accuracy for Imagenet-10 since the ground-truth labels of test samples are invisible
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for the experiments.
Software Dependencies No The paper mentions specific model architectures (Le Net, Res Net-34) but does not provide details on software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes All the hyper-paremters are reported in Appendix Supplemental experiments .