Title: Forgetting is Not Erasure: Recovering Latent Knowledge via Transport Keys

Abstract

Catastrophic forgetting is frequently interpreted as a representational deficit, where models seem to discard the features essential for earlier tasks following sequential training. We contest this rigid interpretation. Through controlled continual-learning experiments, we demonstrate that much of this apparent memory loss stems from interface drift between internal layers rather than the permanent destruction of task-relevant computations. To investigate this, we employ a model stitching protocol that merges early-stage processing from a post-update network with late-stage processing from its predecessor, utilizing compact, task-specific transport keys as optional mediators. We define these transport keys at the systems level as concise interface-alignment operators, derived from limited pairs of anchor activations and validated via model stitching. Our experiments on a ResNet-style architecture using split CIFAR-100 show that transport keys restore the majority of Task A performance after sequential learning of Task B. A comparable recovery trend is observed in a compact vision transformer. These findings imply that continual learning advances should focus not solely on preventing weight updates, but also on developing superior mechanisms for indexing and retrieving latent computations.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC