Work

Replication code and benchmark summaries for: Subspace-Denoised Householder Reflections for Exemplar-Free Continual Linear Probing Under Representation Drift

Year: 2026
Type: software
Abstract: Standalone Python/NumPy/SciPy replication code and benchmark summaries (v1.0.0) accompanying the study “Subspace-Denoised Householder Reflections for Exemplar-Free Continual Linear Probing Under Representation Drift”. Continual linear probing without exemplars leads to catastrophic forgetting on unprobed classes when representations drift across sessions. Updating weights with row-wise gradient, elastic weight consolidation, or ridge methods moves observed classes while leaving unprobed prototypes unchanged in old coordinates. Here we prove that rank-one matrix updates with positive gates have strictly real eigenvalues and cannot perform planar rotations. Signed reflections and even products of Householder matrices instead build orthogonal drift operators with time and memory linear in rank and dimension while preserving prototype Gram structure. We propose Subspace-Householder Continual Probing, which projects anchor prototype shifts onto a dominant drift subspace of rank r in d dimensions. This projection cuts isotropic estimation variance by r/d and shifts all stored class prototypes together. Across 30 multiclass partial-probe drift tasks, we raise balanced accuracy on unprobed old classes from 0.222 to 0.371 and outperform online gradient descent, diagonal elastic weight consolidation, and ridge methods on joint balanced accuracy at p_holm <= 0.0076. Across 25 continual vision and language streams under neural adapter drift, our method outperforms feature-translation linear probing at p_holm = 4.0x10^-5 and continual ridge at p_holm = 0.0276, while trailing 512-unit nonlinear RanPAC at 0.760 versus 0.784.
Language: English
Open Access status: green
OpenAlex ID: W7236664588
Cites:
Cited by:
Related to: