Learning Sample Difficulty from Pre-trained Models for Reliable Prediction

Authors: Peng Cui, Dan Zhang, Zhijie Deng, Yinpeng Dong, Jun Zhu

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

Reproducibility Variable Result LLM Response
Research Type Experimental We perform extensive experiments to verify our method’s effectiveness, showing that the proposed method can improve prediction performance and uncertainty quantification simultaneously.
Researcher Affiliation Collaboration Peng Cui1 5, Dan Zhang2 3, Zhijie Deng4 , Yinpeng Dong1 5, Jun Zhu1 5 1Dept. of Comp. Sci. & Tech., Institute for AI, BNRist Center, THBI Lab, Tsinghua-Bosch Joint ML Center, Tsinghua University, Beijing, 100084 China 2Bosch Center for Artificial Intelligence 3 University of Tübingen 4Qing Yuan Research Institute, Shanghai Jiao Tong University 5Real AI
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any links to open-source code or statements about code availability.
Open Datasets Yes Datasets. CIFAR-10/100 [25], and Image Net1k [7] are used for multi-class classification training and evaluation.
Dataset Splits No The paper mentions "Image Net1k validation sample" and refers to "validation subsets" in Fig. 2 but does not provide specific split percentages or explicitly state the use of predefined standard splits with formal citation.
Hardware Specification No The paper does not specify any hardware used for running the experiments.
Software Dependencies No The paper mentions "SGD with a weight decay of 1e-4 and a momentum of 0.9" which are optimization parameters, not specific software dependencies with version numbers. No other software details are provided.
Experiment Setup Yes Implementation details. We use standard data augmentation (i.e., random horizontal flipping and cropping) and SGD with a weight decay of 1e-4 and a momentum of 0.9 for classification training, and report averaged results from five random runs. The default image classifier architecture is Res Net34 [15]. For the baselines, we use the same hyper-parameter setting as recommended in [52]. For the hyper-parameters in our training loss (9), we set α as 0.3 and 0.2 for CIFARs and Image Net1k, respectively, where T equals 0.7 for all datasets.