UnAct: Gradient-Free Unlearning via Targeted Activation Intervention
UnAct tries to forget a class using only a few target images, without gradients, labels, retained data, or full retraining.
The method scores late-layer units by activation on the forget images, weakens the most responsive connections, and repeats the process for up to 20 rounds. In ResNet-18 tests on CIFAR-10, CIFAR-20, and CIFAR-100, the authors say it stayed competitive with SSD and LFSSD for full-class forgetting and avoided collapse when forget data was scarce. With five CIFAR-10 forget images, UnAct landed 0.21 points from retraining, versus 67 for LFSSD and 90 for SSD. A preliminary ViT-B/16 transfer also favored UnAct, with a reported 19x speedup over SSD when SSD computes importance at request time. HF Daily Papers' note
The method scores late-layer units by activation on the forget images, weakens the most responsive connections, and repeats the process for up to 20 rounds. In ResNet-18 tests on CIFAR-10, CIFAR-20, and CIFAR-100, the authors say it stayed competitive with SSD and LFSSD for full-class forgetting and avoided collapse when forget data was scarce. With five CIFAR-10 forget images, UnAct landed 0.21 points from retraining, versus 67 for LFSSD and 90 for SSD. A preliminary ViT-B/16 transfer also favored UnAct, with a reported 19x speedup over SSD when SSD computes importance at request time. HF Daily Papers' note
score 4