[Paper Review] From Noisy Prediction to True Label: Noisy Prediction Calibration via Generative Model

HeeSun Bae, Seungjae Shin+4 more|arXiv (Cornell University)|May 2, 2022

Machine Learning and Data Classification4 citations

TL;DR

This paper proposes Noisy Prediction Calibration (NPC), a post-processing method that corrects noisy predictions from pre-trained classifiers using a deep generative model to estimate a transition matrix from noisy predictions to true labels. NPC improves classification accuracy across multiple datasets without retraining, outperforming existing methods by significantly reducing prediction errors, especially in detecting previously overlooked noisy instances in benchmark datasets like MNIST and Fashion-MNIST.

ABSTRACT

Noisy labels are inevitable yet problematic in machine learning society. It ruins the generalization of a classifier by making the classifier over-fitted to noisy labels. Existing methods on noisy label have focused on modifying the classifier during the training procedure. It has two potential problems. First, these methods are not applicable to a pre-trained classifier without further access to training. Second, it is not easy to train a classifier and regularize all negative effects from noisy labels, simultaneously. We suggest a new branch of method, Noisy Prediction Calibration (NPC) in learning with noisy labels. Through the introduction and estimation of a new type of transition matrix via generative model, NPC corrects the noisy prediction from the pre-trained classifier to the true label as a post-processing scheme. We prove that NPC theoretically aligns with the transition matrix based methods. Yet, NPC empirically provides more accurate pathway to estimate true label, even without involvement in classifier learning. Also, NPC is applicable to any classifier trained with noisy label methods, if training instances and its predictions are available. Our method, NPC, boosts the classification performances of all baseline models on both synthetic and real-world datasets. The implemented code is available at https://github.com/BaeHeeSun/NPC.

Motivation & Objective

• To address the persistent challenge of noisy labels in machine learning, which degrade classifier generalization and are hard to correct after training.
• To develop a post-processing solution applicable to any pre-trained classifier without access to its training data or architecture.
• To improve the accuracy of noisy predictions by modeling the transition from noisy predictions to true labels using a deep generative model.
• To empirically validate that NPC outperforms existing methods in correcting mispredictions, especially in detecting subtle or previously overlooked label noise.
• To demonstrate the method's robustness across diverse noise types and real-world datasets, including benchmarks assumed to be clean.

Proposed method

• NPC introduces a novel transition matrix estimation method based on a deep generative model that maps noisy predictions to true labels, bypassing the need for true label supervision.
• The method uses a conditional generative model to learn the distribution of inputs given noisy predictions, enabling reverse inference to estimate the true label.
• NPC formulates the calibration as a posterior estimation problem, leveraging the likelihood of inputs given noisy predictions to infer the most probable true label.
• The approach is theoretically grounded in equivalence to traditional transition matrix methods, ensuring consistency with established noise modeling frameworks.
• NPC operates as a plug-in post-processor, requiring only the model’s predictions and input features, making it universally applicable to any black-box classifier.
• It employs GradCAM to visualize attention, showing that NPC focuses on class-relevant features rather than spurious image-level patterns.

Experimental results

Research questions

• RQ1 Can a post-processing method effectively calibrate noisy predictions from a pre-trained classifier without access to training data or architecture?
• RQ2 How does NPC’s generative modeling approach compare to existing transition matrix-based methods in estimating the true label from noisy predictions?
• RQ3 To what extent can NPC detect and correct previously overlooked noisy instances in widely used benchmark datasets like MNIST and Fashion-MNIST?
• RQ4 Does NPC improve generalization performance across diverse noise settings, including symmetric and asymmetric label noise?
• RQ5 How does NPC’s attention mechanism differ from generative models that also generate inputs, and what impact does this have on calibration accuracy?

Key findings

• NPC significantly boosts the accuracy of all baseline models on both synthetic and real-world datasets, with statistical significance, even without access to noise ratios.
• On CIFAR-10 with 20% symmetric noise, NPC improved performance by 5.59× more effective correction (measured by (b)+(f)−((d)+(h))) than the post-processor version of CausalNL.
• NPC successfully identified 50 high-confidence, potentially mislabeled instances in MNIST and Fashion-MNIST, including cases where annotations conflicted with visual content, such as Sneaker vs. Ankle Boot confusion.
• The method demonstrated superior focus on class-relevant features, as shown by GradCAM, whereas CausalNL attended to irrelevant image regions due to its generation objective.
• NPC achieved state-of-the-art performance across all evaluated benchmarks, proving effective even on datasets previously considered clean.
• The Venn diagram analysis confirmed that NPC reduces harmful corrections (case 4) and increases beneficial corrections (case 2), indicating a cautious and reliable calibration strategy.