🤖 AI Summary
In noisy label learning, generative models suffer from reconstruction bias, fixed causal directionality, and neglect of instance-level label uncertainty. To address these issues, we propose a direction-agnostic Expectation-Maximization (EM) framework that eliminates the image generator and instead jointly models the distribution over features, clean labels, and noisy observations. Our method introduces a novel switchable-causal-direction E-step, replacing intractable generative terms with discriminative normalized surrogates, and incorporates an instance-level partial-label prior (PLS) to explicitly capture label uncertainty. This design preserves the statistical rigor of generative modeling while ensuring computational efficiency. Evaluated on benchmark vision and NLP datasets with synthetic and real-world label noise, our approach achieves state-of-the-art accuracy, significantly reduces estimation error of the label transition matrix, and requires substantially less training compute compared to existing generative and discriminative methods.
📝 Abstract
Although noisy-label learning is often approached with discriminative methods for simplicity and speed, generative modeling offers a principled alternative by capturing the joint mechanism that produces features, clean labels, and corrupted observations. However, prior work typically (i) introduces extra latent variables and heavy image generators that bias training toward reconstruction, (ii) fixes a single data-generating direction ((Y
ightarrow!X) or (X
ightarrow!Y)), limiting adaptability, and (iii) assumes a uniform prior over clean labels, ignoring instance-level uncertainty. We propose a single-stage, EM-style framework for generative noisy-label learning that is emph{direction-agnostic} and avoids explicit image synthesis. First, we derive a single Expectation-Maximization (EM) objective whose E-step specializes to either causal orientation without changing the overall optimization. Second, we replace the intractable (p(Xmid Y)) with a dataset-normalized discriminative proxy computed using a discriminative classifier on the finite training set, retaining the structural benefits of generative modeling at much lower cost. Third, we introduce emph{Partial-Label Supervision} (PLS), an instance-specific prior over clean labels that balances coverage and uncertainty, improving data-dependent regularization. Across standard vision and natural language processing (NLP) noisy-label benchmarks, our method achieves state-of-the-art accuracy, lower transition-matrix estimation error, and substantially less training compute than current generative and discriminative baselines. Code: https://github.com/lfb-1/GNL
