An Instance-Dependent Simulation Framework for Learning with Label Noise

TL;DR

This paper introduces an instance-dependent simulation framework for generating realistic noisy labels, enabling better evaluation of learning algorithms and proposing a label correction method that improves performance.

Contribution

The paper presents a novel simulation framework for realistic label noise and a new label correction technique leveraging annotator features.

Findings

Synthetic noisy labels are closer to human labels than traditional methods.

Benchmarking reveals varying algorithm robustness to different noise types.

LQM improves model performance when integrated with existing noisy label techniques.

Abstract

We propose a simulation framework for generating instance-dependent noisy labels via a pseudo-labeling paradigm. We show that the distribution of the synthetic noisy labels generated with our framework is closer to human labels compared to independent and class-conditional random flipping. Equipped with controllable label noise, we study the negative impact of noisy labels across a few practical settings to understand when label noise is more problematic. We also benchmark several existing algorithms for learning with noisy labels and compare their behavior on our synthetic datasets and on the datasets with independent random label noise. Additionally, with the availability of annotator information from our simulation framework, we propose a new technique, Label Quality Model (LQM), that leverages annotator features to predict and correct against noisy labels. We show that by adding LQM as a label correction step before applying existing noisy label techniques, we can further improve the models' performance.