Tom Saliencro, Wen-Hsuan Liu, Rohan Desai, Amara Okonkwo, Maya Lindqvist
Exemplar-free continual learning on pre-trained Vision Transformers increasingly relies on Low-Rank Adaptation (LoRA), yet prevailing methods govern the low-rank update space through strict subspace isolation. By projecting each incoming task into the approximate null space of historical token activations, isolation-based learners equate any cross-task feature overlap with catastrophic interference. We show that this orthogonality premise induces a structural bottleneck we term \emph{spectral starvation}: when down-projection bases are locked before task optimization, the gradient norm reaching the trainable up-projection is upper-bounded by the energy of current-task features captured by the chosen basis. Because natural visual tasks share low- and mid-level covariance structure, historical null tails contain barely $5.5\%$ of incoming task variance, starving new-class plasticity while blocking positive consolidation along shared directions. We propose BISECT (Bilevel Spectral Decomposition), a continual LoRA framework that partitions a fixed rank-$r$ budget into two orthogonal functional blocks: an invariant carrier spanned by the dominant modes of the joint historical-and-current Gram matrix, and a novelty complement extracted via generalized Rayleigh whitening followed by orthogonal complement projection. After single-task training, BISECT folds both blocks back into the frozen backbone using a closed-form precision gate that scales each rank-1 invariant atom by its current-to-joint Rayleigh ratio while preserving novelty atoms at unit gain. Across eight incremental settings on ImageNet-R, ImageNet-A, CIFAR-100, CUB-200, and DomainNet, BISECT outperforms prompt, adapter, and LoRA baselines with zero extra inference parameters or FLOPs, and remains complementary to sparse analytic expansion.
Wen-Hsuan Liu, Amara Okonkwo, Tom Saliencro, Rohan Desai, Maya Lindqvist
Parameter-efficient continual learning on pre-trained Vision Transformers largely relies on a costly premise: preventing catastrophic forgetting requires either spawning a dedicated LoRA branch per session---causing adapter parameters to scale linearly with task horizon---or applying sparse Mixture-of-Experts gating symmetrically to both forward and backward passes, which sacrifices representational rank at test time. We argue that reading consolidated knowledge and writing new task credit are fundamentally asymmetric operations. Because visual features in a shared low-rank space distribute across rank-1 atoms rather than localizing in disjoint task silos, forward inference should read all atoms at full rank capacity, whereas backward optimization should restrict plasticity to a sparse subset of input-aligned channels. We introduce RASP (Read-All, Sparse-Plasticity LoRA), which adapts a single stationary-size adapter across sequential tasks. At each step, a scene-modulated cosine router selects the top-$q$ unlocked atoms to receive gradients, blocks updates on the remaining $r-q$ channels, and penalizes Frobenius Gram coherence between active and non-updated output directions. Across sessions, the router's empirical activation trace and each atom's exact spectral norm form a zero-overhead geometric consolidation index $ω_j = \sqrt{\tilde{c}_j \tildeσ_j}$ that drives Cauchy step attenuation and boundary anchor locking. Across CIFAR-100, CUB-200, ImageNet-R, and ImageNet-A, RASP matches or surpasses branch-expanding LoRA learners under a strictly constant parameter footprint.