Learning from Training Dynamics: Identifying Mislabeled Data beyond Manually Designed Features

Authors: Qingrui Jia, Xuhong Li, Lei Yu, Jiang Bian, Penghao Zhao, Shupeng Li, Haoyi Xiong, Dejing Dou

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments to evaluate the proposed method. We train the noise detector based on the synthesized label-noised CIFAR dataset and test such noise detector on Tiny Image Net, CUB-200, Caltech-256, Web Vision and Clothing1M. Results show that the proposed method precisely detects mislabeled samples on various datasets without further adaptation, and outperforms state-of-the-art methods.
Researcher Affiliation Collaboration 1 Sino-French Engineer School, Beihang University, Beijing, China 2 Baidu Inc. Beijing, China 3 Beihang Hangzhou Innovation Institute Yuhang, Hangzhou, China 4 BCG Greater China, Beijing, China
Pseudocode Yes Algorithm 1: Supervised Learning from Training Dynamics.
Open Source Code Yes Code available at https://github.com/Christophe-Jia/mislabeldetection and at Interpret DL (Li et al. 2022) as well.
Open Datasets Yes Evaluations are conducted on four datasets with synthesized noisy labels: CIFAR10/100 (Krizhevsky and Hinton 2009), CUB-200-2011 (Wah et al. 2011) and Caltech256 (Griffin, Holub, and Perona 2007). ...change the labels of the training set with symmetry (or asymmetry) noises but leave the testing set unchanged.
Dataset Splits No The paper mentions using 'training set' and 'testing set' for experiments but does not explicitly specify a 'validation set' or its split.
Hardware Specification No The paper mentions using Res Nets and LSTM models but does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running experiments.
Software Dependencies No The paper mentions using Res Nets and an Adam W optimizer but does not specify software libraries with version numbers (e.g., PyTorch 1.x, TensorFlow 2.x, Python 3.x) that were used.
Experiment Setup Yes Experiment Setups. Evaluations are conducted on four datasets with synthesized noisy labels: CIFAR10/100 (Krizhevsky and Hinton 2009), CUB-200-2011 (Wah et al. 2011) and Caltech256 (Griffin, Holub, and Perona 2007). ...We therefore choose Res Nets as the deep model f ... an LSTM of two hidden layers with 64 dimensions for each is used. ...The optimization process is guided by an Adam W optimizer...