Conditional Variational Autoencoder Data Augmentation for Robustness to Instance-Dependent Noise

Research Paper

For detailed information about our study and findings, please refer to our research paper. You can access it here: Report Paper

Abstract

This paper addresses the critical challenge in deep learning of obtaining high-quality, accurately labeled data. In real-world scenarios, label noise is often unavoidable and can significantly hinder model performance. Traditional artificially generated label noise does not adequately represent real-world conditions. We propose a novel approach using instance-dependent noise (IDN) as a benchmark for modeling real-world label noise, coupled with a deep generative method for effective handling of this issue.

Introduction

The effectiveness of deep learning systems heavily relies on the quality of the labeled data. However, accurate and noise-free labels are challenging and time-consuming to obtain. Our research focuses on addressing the gap between artificially generated label noise and real-world label noise, using IDN as a more representative benchmark.

Methodology

We introduce the Conditional Variational Autoencoder (CVAE) as a novel method for data augmentation on desired class labels. This approach aims to enhance the robustness of deep learning models against the detrimental effects of label noise.

Conditional Variational Autoencoder (CVAE)

• The CVAE is designed to encode generalized class-conditional features.
• It helps in mitigating the influence of noisy labels by enhancing data quality through augmentation.

Dataset and Evaluation

• Dataset Used: MNIST dataset.
• Evaluation Method: Comparative analysis with several state-of-the-art techniques under less noisy conditions.

Results

• The experimental results demonstrate that our CVAE-based method outperforms existing techniques in scenarios with lower levels of noise.
• The CVAE-based data augmentation shows robustness to small percentages of noise, underscoring its potential in real-world image classification scenarios.

Conclusion

The study confirms the potential of the CVAE in improving the performance of image classification models, particularly in real-world settings where label noise is a common issue. Our findings highlight the importance of developing techniques that are tailored to the nuances of real-world data.