MAAT: Multi-phase Adapter-Aware Targeted Unlearning
Researchers have introduced MAAT, a novel three-phase framework for targeted machine unlearning that specifically addresses the difficulty of removing causal knowledge. Existing benchmarks are skewed, underrepresenting "Why" questions, which are crucial for evaluating causal and relational knowledge removal. MAAT operates on LoRA adapter weights and employs techniques like gradient-projected ascent and SVD pruning to achieve high forgetting while retaining other knowledge. The accompanying 5WBENCH benchmark, with balanced categories for Who, What, When, Where, and Why, quantifies these unlearning failures for the first time. AI
IMPACT Introduces a new benchmark and method to improve the evaluation and execution of machine unlearning, particularly for causal knowledge.