Adaptive estimation of instance-dependent noise transition matrix for learning with instance-dependent label noise.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Instance-dependent noise (IDN) is a significant challenge in real-world datasets, impacting deep neural network performance.
• Existing methods for IDN often rely on estimating the instance-dependent noise transition matrix (IDNT) with restrictive assumptions or anchor points, leading to estimation errors.

Purpose of the Study:

• To develop a novel method for estimating the instance-dependent noise transition matrix (IDNT) without making assumptions about its form or requiring anchor points.
• To improve the robustness of deep neural networks in the presence of instance-dependent noise.

Main Methods:

• Computed instance-label confusion matrix (ILCM) using similarity scores between instance features and label representations to capture noise characteristics.
• Adaptively combined noisy class posteriors with ILCM, weighted by noise degree, for enhanced IDNT estimation.
• Adjusted the loss function using the estimated IDNT to train a more robust classifier.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art approaches on synthetic datasets (F-MNIST, SVHN, CIFAR-10, CIFAR-100).
• Effectiveness was validated on a real-world noisy dataset (Clothing1M).
• The approach successfully mitigates the impact of instance-dependent noise, leading to improved classification accuracy.

Conclusions:

• The developed method effectively estimates IDNT without prior assumptions, offering a more flexible and accurate approach to handling instance-dependent noise.
• This technique enhances deep neural network robustness and performance on datasets with complex noise patterns.
• The findings suggest a promising direction for developing more resilient machine learning models for real-world applications.