Cost-Sensitive Self-Training for Optimizing Non-Decomposable Metrics

Authors: Harsh Rangwani, shrinivas ramasubramanian, Sho Takemori, Kato Takashi, Yuhei Umeda, Venkatesh Babu R

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our results demonstrate that CSST achieves an improvement over the state-of-the-art in majority of the cases across datasets and objectives. We demonstrate that the proposed CSST framework shows significant gains in performance on both vision and NLP tasks on imbalanced datasets, with an imbalance ratio defined on the training set as ρ = maxi P (y=i) mini P (y=i) .
Researcher Affiliation Collaboration 1Video Analytics Lab, Indian Institute of Science, Bengaluru, India 2Fujitsu Limited, Kanagawa, Japan
Pseudocode Yes See Alg. 1 in Appendix
Open Source Code Yes Code: https://github.com/val-iisc/Cost Sensitive Self Training
Open Datasets Yes For CIFAR-10 [11], IMDb [16] and DBpedia-14 [14], we use ρ = 100 and ρ = 10 for CIFAR-100 [11] and Image Net-100 [27] 2 datasets.
Dataset Splits Yes We divide the balanced held-out set for each dataset equally into validation and test sets.
Hardware Specification No The paper states 'See Section G in Appendix' for details on hardware specification, but these details are not provided in the main body of the paper.
Software Dependencies No The paper mentions using Wide Res Nets(WRN), Res Net-50, and Distil BERT, but does not provide specific version numbers for software dependencies or libraries.
Experiment Setup No The paper states 'A detailed list of hyper-parameters and additional experiments can be found in the Appendix Tab. 4 and Sec. O respectively', but these details are not explicitly provided in the main text.