On the Plasticity Collapse in Continual Machine Unlearning
Sequential unlearning may make later forgetting worse and even bring back data the model had already “forgotten.”
The paper names this failure “plasticity collapse”: repeated unlearning requests pile up constraints in parameter space until future updates are restricted. The authors describe two failure modes, with later tasks showing weaker forgetting and earlier removed information reappearing. Experiments across image-classification architectures, datasets, and unlearning methods are presented as evidence that the effect is not implementation-specific. ArXiv · AI/CL/LG's note
The paper names this failure “plasticity collapse”: repeated unlearning requests pile up constraints in parameter space until future updates are restricted. The authors describe two failure modes, with later tasks showing weaker forgetting and earlier removed information reappearing. Experiments across image-classification architectures, datasets, and unlearning methods are presented as evidence that the effect is not implementation-specific. ArXiv · AI/CL/LG's note
score 5