UnAct: Gradient-Free Unlearning Matches Retraining With Just 5 Forget Images

GAASH-Lab · hf · 2026-10-07

GAASH-Lab introduces UnAct, a gradient-free class-unlearning method that requires only forward passes over forget images.

How it works

• Unlike SSD/LFSSD, which still need backpropagation over the full dataset, UnAct scores late-layer units by their responses and attenuates the most responsive connections over up to 20 rounds
• No gradients, labels, or retained data needed

Results (ResNet-18)

• Competitive with SSD/LFSSD on full-class forgetting but never collapses the network when forget data is scarce
• Retain accuracy stays within 2.5 points of retraining, while SSD/LFSSD lose up to 86 points on some classes
• With 5 forget images on CIFAR-10: 0.21-point distance to retraining vs 67 for LFSSD and 90 for SSD
• On ViT-B/16: 11.5 vs SSD's 33.7, and 19x faster

Code: github.com/abdulmuizz0903/UnAct

Original post →

More from Research

• Converting 10 coding harnesses including Claude Code into RL environments, cutting tool calls 31% — vllm_project · 2026-10-07
• Google DeepMind's IdeaLens reads outlines to detect AI-generated ideas, not AI-written words — rohanpaul_ai · 2026-10-07
• Anima Anandkumar's First UN Visit: The 'Silent Revolution' of AI in Science — AnimaAnandkumar · 2026-10-07
• WRAP: Adversarial Training Tackles Deep Hedging in Nonstationary Markets via DRO — chaumian · 2026-10-07
• NYU's CIA Metric Finds LLMs' Chain-of-Thought Only 44.8-75.9% Aligned With Internal Computation — newyorkuniversity · 2026-10-07
• Oxford's AI Theorist Autonomously Derives Physical Model for Quantum Material α-RuCl₃ — Oxford · 2026-10-07

Research channel →