FINE Samples for Learning with Noisy Labels

Authors: Taehyeon Kim, Jongwoo Ko, sangwook Cho, JinHwan Choi, Se-Young Yun

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that the proposed methods consistently outperform corresponding baselines for all three applications on various benchmark datasets. 1 Introduction Deep neural networks (DNNs) have achieved remarkable success in numerous tasks as the amount of accessible data has dramatically increased [21, 15].
Researcher Affiliation Academia Taehyeon Kim KAIST AI KAIST Daejeon, South Korea potter32@kaist.ac.kr Jongwoo Ko KAIST AI KAIST Daejeon, South Korea jongwoo.ko@kaist.ac.kr Sangwook Cho KAIST AI KAIST Daejeon, South Korea sangwookcho@kaist.ac.kr Jinhwan Choi KAIST AI KAIST Daejeon, South Korea jinhwanchoi@kaist.ac.kr Se-Young Yun KAIST AI KAIST Daejeon, South Korea yunseyoung@kaist.ac.kr
Pseudocode Yes Algorithm 1: FINE Algorithm for Sample Selection
Open Source Code Yes Code available at https://github.com/Kthyeon/FINE_official
Open Datasets Yes Noisy Benchmark Dataset. Following the previous setups [25, 29], we artificially generate two types of random noisy labels: injecting uniform randomness into a fraction of labels (symmetric) and corrupting a label only to a specific class (asymmetric). For example, we generate noise by mapping TRUCK AUTOMOBILE, BIRD AIRPLANE, DEER HORSE, CAT DOG to make asymmetric noise for CIFAR-10. For CIFAR-100, we create 20 five-size super-classes and generate asymmetric noise by changing each class to the next class within super-classes circularly. For a real-world dataset, Clothing1M [44] containing inherent noisy labels is used.
Dataset Splits Yes The dataset provides 50k, 14k, and 10k verified as clean data for training, validation, and testing.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory) are mentioned for the experiments.
Software Dependencies No No specific ancillary software details with version numbers (e.g., PyTorch 1.9, Python 3.8) are provided.
Experiment Setup Yes We set the threshold ζ as 0.5.