Networks hide erased features instead of fully deleting them
Hidden not Deleted: How Networks Suppress Entangled Features
Machine LearningArtificial Intelligence
Summary
When trying to erase certain features in artificial neural networks by isolating them in separate math spaces, this approach can fail if features overlap in complex ways. The authors show that networks trained with typical methods find two different ways to hide these features instead of eliminating them. This hiding means the features can often be recovered later, explaining why some 'unlearning' attempts in language models don’t fully forget information. Their work provides a clear explanation of why erased knowledge sometimes still lurks inside networks.
What this means in practice
- •For machine learning engineers: Identify whether unlearning methods truly delete unwanted features or merely suppress them to avoid later reactivation.
- •For ai safety teams: Improve verification protocols to confirm that sensitive or harmful information is thoroughly removed from trained models.
Authors
Akash Samanta, Manish Pratap Singh, Debasis Chaudhuri
Abstract
Concept erasure methods that operate via linear projection assume that features occupy separable subspaces. We show this assumption fails under dense superposition: when two features are forced into an antipodal pair sharing a single subspace, state-of-the-art linear erasure destroys both, not just the target. Networks trained with gradient descent instead solve this problem non-linearly, but not uniformly: they converge to one of two distinct circuit-level solutions depending on initialization, which we call mirror and shadow solutions. We map this bifurcation as a function of feature entanglement, show it reflects a stable attractor structure rather than an artifact of our setup, and use targeted causal interventions to demonstrate that both solutions leave a substantial, measurable trace of the erased feature's representation intact, recoverable through a single scalar patch rather than requiring any further training. This mirrors a failure mode recently observed empirically in LLM unlearning, where suppression rather than deletion allows forgotten knowledge to resurface; our results offer a mechanistic, causally-validated account of why that failure mode occurs.