PureGen purifies poisoned training data using iterative Langevin dynamics of Energy-Based Models and Denoising Diffusion Probabilistic Models, achieving state-of-the-art poison defense with minimal impact on classifier generalization.
A composable latent space augmentation framework using VAEs that allows augmentations to be combined through linear transformations, preserving specific augmentation variances and improving geometric interpretability over standard and Conditional VAEs.
CCGM combines a causal latent space VAE with modifications for causal fidelity to generate de-biased datasets from biased training data, offering fine-grained control over causal structure in both image and tabular data generation.