Data Unlearning via Inverse Distillation
IDU trains a one-step generator while pushing out a specified forget set, without needing the retained training data.
The paper frames unlearning and distillation as one objective for flow and diffusion-style models. It uses a pretrained full-data teacher plus examples from the forget set, avoiding extra classifiers, feature extractors, or retained examples. In MNIST and CIFAR-10 tests, the method reduced generation of forgotten classes while keeping retained-class quality. HF Daily Papers' note
The paper frames unlearning and distillation as one objective for flow and diffusion-style models. It uses a pretrained full-data teacher plus examples from the forget set, avoiding extra classifiers, feature extractors, or retained examples. In MNIST and CIFAR-10 tests, the method reduced generation of forgotten classes while keeping retained-class quality. HF Daily Papers' note
score 5