Paper accepted to ICLR26 conference! [Dataset Distillation as Pushforward Optimal Quantization]

Our work on dataset distillation has been accepeted to the ICLR’26 conference! Dataset distillation consists of reducing the size of a training dataset, such that training on the reduced dataset gives a model with similar performance to the whole dataset. We interpret a new paradigm of dataset distillation, namely “disentangled” methods, as actually performing clustering inside a latent space. In particular, under the common diffusion model encoder-decoder backbones, disentangled methods exactly perform clustering, allowing us to translate the classical convergence rates to the dataset setting using Wasserstein-L1 duality. openreview