Disentangling Sampling and Labeling Bias for Learning in Large-output Spaces

Authors: Ankit Singh Rawat, Aditya K Menon, Wittawat Jitkrittum, Sadeep Jayasumana, Felix Yu, Sashank Reddi, Sanjiv Kumar

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically verify our findings on long-tail classification and retrieval benchmarks. and 5. Experiments We now present experiments on benchmarks for both long-tail learning and retrieval, illustrating our main finding: existing negative sampling schemes, such as within-batch sampling with constant weighting, implicitly trade-off performance on dominant versus rare labels.
Researcher Affiliation Industry 1Google Research, New York, USA.
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide an explicit statement or link to the source code for the described methodology.
Open Datasets Yes We present results on long-tailed ( LT ) versions of the CIFAR-100 and Image Net datasets. and In particular, we experiment with AMAZONCAT-13K and WIKILSHTC-325K datasets from the extreme classification literature (Agrawal et al., 2013; Bengio et al., 2019), where due to a large number of labels it is common to employ negative sampling. In addition, we also explored a small scale dataset DELICIOUS from the repository to make our conclusions more general.
Dataset Splits No The paper mentions training on datasets and evaluating on a test set, but does not explicitly provide the training/validation/test split percentages or counts for all datasets in the main text. It states 'We report the test set balanced error,' which indicates a test set, but a clear validation split is not specified.
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments, such as GPU or CPU models.
Software Dependencies No The paper mentions training models (e.g., ResNet) but does not provide specific version numbers for software dependencies or libraries used.
Experiment Setup Yes We use m = 32 negatives on CIFAR-100, and m = 512 negatives on Image Net. and We train a Res Net-56 for CIFAR and a Res Net-50 for Image Net, using SGD with momentum; see Appendix E for details.