Maximising the Utility of Validation Sets for Imbalanced Noisy-label Meta-learning

Authors: Hoang Anh Dung, Cuong C. Nguyen, Vasileios Belagiannis, Thanh-Toan Do, Gustavo Carneiro

TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The proposed method shows state-of-the-art (SOTA) results compared to previous meta-learning and noisy-label learning approaches on several noisy-label learning benchmarks. 5 Experiments The proposed method INOLML is evaluated on several datasets, including CIFAR10, CIFAR100, mini Web Vision and Controlled Noisy Web Labels (CNWL).
Researcher Affiliation Academia Dung Anh Hoang EMAIL Department of Data Science and AI Monash University Cuong Nguyen EMAIL Centre for Vision, Speech and Signal Processing University of Surrey Vasileios Belagiannis EMAIL Friedrich-Alexander-Universität Erlangen-Nürnberg Thanh-Toan Do EMAIL Department of Data Science and AI Monash University G. Carneiro EMAIL Centre for Vision, Speech and Signal Processing University of Surrey
Pseudocode Yes Algorithm 1 Training procedure of the proposed INOLML.
Open Source Code No The paper does not provide concrete access to source code for the methodology described in this paper.
Open Datasets Yes The proposed method INOLML is evaluated on several datasets, including CIFAR10, CIFAR100, mini Web Vision and Controlled Noisy Web Labels (CNWL). Both CIFAR10 and CIFAR100 datasets (Krizhevsky & Hinton, 2009) contain 50k and 10k images used for training and testing, respectively. Web Vision (Li et al., 2017) is a dataset of 2.4 million images crawled from Google and Flickr...
Dataset Splits Yes Both CIFAR10 and CIFAR100 datasets (Krizhevsky & Hinton, 2009) contain 50k and 10k images used for training and testing, respectively. Red mini-Image Net dataset consisting of 50k training images from 100 classes for training and 5k images for testing. For instance, we need around K = 200 candidate samples per class for D(v) before selecting M = 10 samples per class for the validation set D(v) (e.g., for CIFAR100, we need 20,000 samples for D(v))
Hardware Specification Yes The experiment above was conducted on a single NVIDIA V100 GPU. For the Web Vision experiment, we use p = 4, k = 8 with 4 NVIDIA V100 GPU and batches of size 16.
Software Dependencies No The paper does not provide specific software dependencies with version numbers.
Experiment Setup Yes For CIFAR, we train Wide Resnet28-10 with 100k iterations and a batch size of 100. For mini-Web Vision, we follow FSR (Zhang & Pfister, 2021) and train a single Resnet50 network with 1 million iterations and a batch size of 16. All experiments use N = 200, K = 50, κ = 0.9.