Tackling Instance-Dependent Label Noise via a Universal Probabilistic Model

Authors: Qizhou Wang, Bo Han, Tongliang Liu, Gang Niu, Jian Yang, Chen Gong10183-10191

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on datasets with both synthetic and real-world label noise verify that the proposed method yields significant improvements on robustness over state-of-the-art counterparts.
Researcher Affiliation Academia 1 Key Laboratory of Intelligent Perception and Systems for High-Dimensional Information of Mo E, School of Computer Science and Engineering, Nanjing University of Science and Technology 2 Department of Computer Science, Hong Kong Baptist University 3 Trustworthy Machine Learning Lab, School of Computer Science, Faculty of Engineering, The University of Sydney 4 RIKEN Center for Advanced Intelligence Project (AIP) 5 Department of Computing, Hong Kong Polytechnic University
Pseudocode Yes Algorithm 1 The Overall Algorithm.
Open Source Code No The paper does not provide any link to open-source code or explicitly state that code is available.
Open Datasets Yes We first test the proposed method on CIFAR-100 and CIFAR10 (Krizhevsky, Hinton et al. 2009) with synthetic label noise, and then conduct real-world label noise experiments on Clothing1M (Xiao et al. 2015).
Dataset Splits Yes Clothing1M: ...it also contains 50k, 14k, and 10k correctly labeled instances for auxiliary training, validation, and test.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions models like 'Res Net-32' and 'Res Net-50' but does not provide specific software dependencies with version numbers (e.g., Python, TensorFlow, PyTorch versions).
Experiment Setup Yes Meansubtraction, horizontal random flip, and 32 32 random crop are performed for data pre-processing. Then, we use minibatch gradient ascent with a momentum of 0.9; a weight decay of 10 4; and a batch size of 256. The alternating optimization algorithm is executed for 160 epochs. For the classifier parameters, we begin with a learning rate α2 = 0.05 and divide it by 10 per 40 epochs. For label confusing probabilities, they are updated every 5 epochs from the 35th epoch. We test model performance with different learning rate α1 on CIFAR-100, and fix α1 = 0.7 on CIFAR-10.