Making Machine Learning Robust to Noisy Labels: Our Latest Work in TMLR

04/29/2025

The increasing availability of pre-trained foundation models has opened new research directions for building efficient and explainable machine learning systems. In our latest paper, An Embedding is Worth a Thousand Noisy Labels, published in Transactions on Machine Learning Research (TMLR), we present a lightweight yet powerful method for mitigating the harmful effects of label noise using a novel embedding-based strategy.

Weighted Adaptive Nearest Neighbour (WANN)

We propose WANN (Weighted Adaptive Nearest Neighbor), a training-free method that utilizes image embeddings from foundation models and a weighted k-NN approach to enhance classification robustness against noisy labels. By introducing a reliability score to assess training label correctness, WANN adaptively selects the neighborhood size for each test sample, weighting predictions based on this score. This improves classification performance over traditional methods like robust loss functions and fixed k-NNs. Furthermore, WANN is lightweight, interpretable, privacy-friendly, generalizing effectively across various datasets, noise types, and domains such as healthcare.

About TMLR

Transactions on Machine Learning Research (TMLR) is a new venue for dissemination of machine learning research that is intended to complement JMLR while supporting the unmet needs of a growing ML community.