Noisy label learning has been tackled with both discriminative and generative approaches. Despite the simplicity and efficiency of discriminative methods, generative models offer a more principled way of disentangling clean and noisy labels and estimating the label transition matrix. However, existing generative methods often require inferring additional latent variables through costly generative modules or heuristic assumptions, which hinder adaptive optimisation for different causal directions. They also assume a uniform clean label prior, which does not reflect the sample-wise clean label distribution and uncertainty. In this paper, we propose a novel framework for generative noisy label learning that addresses these challenges. First, we propose a new single-stage optimisation that directly approximates image generation by a discriminative classifier output. This approximation significantly reduces the computation cost of image generation, preserves the generative modelling benefits, and enables our framework to be agnostic in regards to different causality scenarios (i.e., image generate label or vice-versa). Second, we introduce a new Partial Label Supervision (PLS) for noisy label learning that accounts for both clean label coverage and uncertainty. The supervision of PLS does not merely aim at minimising loss, but seeks to capture the underlying sample-wise clean label distribution and uncertainty. Extensive experiments on computer vision and natural language processing (NLP) benchmarks demonstrate that our generative modelling achieves state-of-the-art results while significantly reducing the computation cost. Our code is available at https://github.com/lfb-1/GNL.

翻译：噪声标签学习已通过判别式与生成式方法得到广泛研究。尽管判别式方法简单高效，但生成式模型为解耦干净标签与噪声标签、估计标签转移矩阵提供了更规范的理论框架。然而，现有生成式方法通常需要借助高成本的生成模块或启发式假设来推断额外隐变量，导致难以针对不同因果方向进行自适应优化。此外，这些方法假设统一先验的干净标签分布，无法反映样本级别的干净标签分布与不确定性。本文提出一种新型生成式噪声标签学习框架以应对上述挑战。首先，我们提出一种新型单阶段优化方法，通过判别式分类器输出直接近似图像生成过程。该近似显著降低了图像生成的计算成本，同时保留生成式建模的优势，使框架能够适用于不同因果场景（即图像生成标签或反之）。其次，我们引入一种新的部分标签监督用于噪声标签学习，该机制同时兼顾干净标签覆盖范围与不确定性。部分标签监督的目标不仅是损失最小化，更旨在捕获样本级别的底层干净标签分布与不确定性。在计算机视觉与自然语言处理基准上的大量实验表明，我们的生成式建模在显著降低计算成本的同时实现了最先进的性能。代码开源于 https://github.com/lfb-1/GNL。