Is Seeing Still Not Necessarily Believing?
By Siyu An, Yating Chen, JangHyeon Lee, Sidharth Singh
Despite remarkable strides in computer vision, the data-hungry nature of deep neural networks poses challenges, particularly in specialized fields like medicine, where data is often scarce. Previous efforts utilized BigGANs for data augmentation, but their photorealistic samples didn't serve as useful additional training data, raising the question, "Is Seeing Necessarily Believing?" In this study, we replace BigGANs with diffusion models to generate synthetic versions of original datasets for downstream classification tasks. Our findings indicate that these diffusion models can create synthetic data that enhance model performance and even substitute original data without significant performance loss, suggesting that seeing can indeed be believing.
