Meta Label Correction for Noisy Label Learning

Leveraging weak or noisy supervision for building effective machine learning models has long been an important research problem. Its importance further increased recently due to the growing need large-scale datasets train deep models. Weak could originate from multiple sources including non-expert annotators automatic labeling based on heuristics user interaction signals. There is extensive amount of previous work focusing leveraging labels. Most notably, recent shown impressive gains by using a meta-learned instance re-weighting approach where meta-learning framework used assign weights In this paper, we extend via posing problem as label correction within framework. We view procedure meta-process and propose new termed MLC (Meta Label Correction) with Specifically, network adopted meta-model produce corrected labels while main model trained leverage Both are jointly solving bi-level optimization run experiments different noise levels types both image recognition text classification tasks. compare re-weighing approaches showing that framing addresses some limitations re-weighting. also show proposed outperforms methods in language