Conditional Diffusion with Label Smoothing for Data Synthesis from Examples with Noisy Labels

Gentry Atkinson, Xiaomin Li, Vangelis Metsis

Abstract: Data generation techniques are critical in various fields where obtaining real-world data is difficult or expensive. However, data generators like generative adversarial networks (GANs) and diffusion-based models depend on training models using pre-existing data and labels. When the labels are unreliable, the performance of data generators can suffer. This paper proposes a novel adaptation of Denoising Diffusion Probabilistic Models (DDPM) that employs label smoothing to enhance the reliability of the generated data in the presence of label noise. Label smoothing mitigates the impact of label noise by preventing the model from becoming overconfident in mislabeled instances of data. We demonstrate that DDPM with label smoothing outperforms both conditional and unconditional DDPM in terms of the closeness of the generated data to the original data's distribution, even when the training data contains instances with mislabeled labels.

External IDs:dblp:conf/eusipco/AtkinsonLM23