Abstract:Noise transition matrix estimation is a promising approach for learning with label noise. It can infer clean posterior probabilities, known as Label Distribution (LD), based on noisy ones and reduce the impact of noisy labels. However, this estimation is challenging, since the ground truth labels are not always available. Most existing methods estimate a global noise transition matrix using either correctly labeled samples (anchor points) or detected reliable samples (pseudo anchor points). These methods heavily rely on the existence of anchor points or the quality of pseudo ones, and the global noise transition matrix can hardly provide accurate label transition information for each sample, since the label noise in real applications is mostly instance-dependent. To address these challenges, we propose an Instance-dependent Label Distribution Estimation (ILDE) method to learn from noisy labels for image classification. The method's workflow has three major steps. First, we estimate each sample's noisy posterior probability, supervised by noisy labels. Second, since mislabeling probability closely correlates with inter-class correlation, we compute the inter-class correlation matrix to estimate the noise transition matrix, bypassing the need for (pseudo) anchor points. Moreover, for a precise approximation of the instance-dependent noise transition matrix, we calculate the inter-class correlation matrix using only mini-batch samples rather than the entire training dataset. Third, we transform the noisy posterior probability into instance-dependent LD by multiplying it with the estimated noise transition matrix, using the resulting LD for enhanced supervision to prevent DCNNs from memorizing noisy labels. The proposed ILDE method has been evaluated against several state-of-the-art methods on two synthetic and three real-world noisy datasets. Our results indicate that the proposed ILDE method outperforms all competing methods, no matter whether the noise is synthetic or real noise.

On the Noise Estimation Statistics

Image noise level estimation based on affine reconstruction and noise sample histogram

Estimating Per-Class Statistics for Label Noise Learning

Noise is the Fatal Poison: A Noise-aware Network for Noisy Dataset Classification

A Label Noise Robust Stacked Auto-Encoder Algorithm for Inaccurate Supervised Classification Problems

Robust Estimation for Image Noise Based on Eigenvalue Distributions of Large Sample Covariance Matrices.

Spectral-Spatial Joint Noise Estimation for Hyperspectral Images

A Novel Approach of Noise Statistics Estimate Using H ∞ Filter in Target Tracking.

Class-Wise Denoising for Robust Learning under Label Noise

Back-End: A Noise Rate Estimation Method In The Presence Of Class Conditional Noise

Instance-dependent Label Distribution Estimation for Learning with Label Noise

Label noise detection under the Noise at Random model with ensemble filters

NoiseRank: Unsupervised Label Noise Reduction with Dependence Models

Machine-Learning-Based Source Number Estimation Under Unknown Spatially Correlated Noise

Tackling Instance-Dependent Label Noise with Dynamic Distribution Calibration

Adaptive Iterative Nonparametric Spectral Estimation Method Based on Generalized Noise Covariance

Uncertainty-Aware Learning against Label Noise on Imbalanced Datasets

A Bi-level Formulation for Label Noise Learning with Spectral Cluster Discovery

On the Robustness of Nearest Neighbor with Noisy Data

Error-Bounded Correction of Noisy Labels

Estimating Noisy Class Posterior with Part-level Labels for Noisy Label Learning