paper

The Stable Recovery Manifold: Geometric Principles Governing Recoverability in Continual Learning

paperactiveprovisionalthe-stable-recovery-manifold-geometric-principles-governing-recoverability-in-continual-learning-3670eef9·1 events·first seen 5d ago

Aliases: The Stable Recovery Manifold: Geometric Principles Governing Recoverability in Continual Learning

Co-occurring entities

Split CIFAR-100 Recovery Subspace Dimensionality ResNet-18

More like this (12)

Recoverable but Not Stationary: Local Linear Structures in Weights and Activations Preserving Plasticity in Continual Learning via Dynamical Isometry TailLoR: Protecting Principal Components in Parameter-Efficient Continual Learning Latent World Recovery for Multimodal Learning with Missing Modalities Recovery Subspace Dimensionality Continual Learning Provenance-Grounded Gating and Adaptive Recovery in Synthetic Post-Training Data Curation Sparse Subspace-to-Expert Sharing for Task-Agnostic Continual Learning Latent World Recovery Backdoor Unlearning Generalization: A Path Toward the Removal of Unknown Triggers in LLMs Continual LLM Upcycling: A Predictor-Gated Bank-Wise Sparsity Training Recipe for Dense-to-Sparse LLMs Learning from the Self-future: On-policy Self-distillation for dLLMs

Recent events (1)

5arXiv · cs.LG·5d ago·source ↗

Stable Recovery Manifold hypothesis: catastrophic forgetting as accessibility problem, not information destruction

A new arXiv preprint investigates the geometric structure of recoverability in continual learning using Split CIFAR-100 and a sequentially trained ResNet-18. The authors introduce Recovery Subspace Dimensionality (k_t) and find that recovery dimensionality remains stable across tasks (mean k_t = 8.0) despite substantial representational drift, with principal-angle drift strongly predicting recoverability (r = -0.862). The findings support the Stable Recovery Manifold hypothesis: forgotten knowledge remains compactly decodable, reframing catastrophic forgetting as a manifold-alignment and accessibility problem rather than true information loss.

Evaluation and Benchmarking Split CIFAR-100 Recovery Subspace Dimensionality The Stable Recovery Manifold: Geometric Principles Governing Recoverability in Continual Learning +1 more