Attacking Noisy Labels Via Dirichlet Process Mixture Models
Yu Song, Zhengping Ren, Si–yi Dai, Jian–wei Liu · 2025
In this paper, we present a novel approach to addressing the challenges posed by noisy labels in classification tasks. The Gaussian Mixture Discriminant Analysis (GMDA) method demonstrates strong performance on both synthetic and real-world datasets affected by label noise. However, GMDA requires manual adjustment of the number of mixture components, a critical parameter that, if improperly set, can negatively impact computational efficiency and prediction accuracy. To tackle this issue, we propose an adaptive model that dynamically determines the optimal number of cluster components, thereby enhancing the classification performance on datasets with noisy labels. We employ the Dirichlet process mixture model, a non-parametric Bayesian method known for its flexibility in handling an unknown number of components. By integrating a flipping probability mechanism, our model effectively learns from datasets corrupted by noise, representing a necessary and significant advancement in the field of noisy label classification.