Adaptive repartitioning of the spectral core

Investigate adaptive repartitioning strategies for the SPARCL spectral core and residual subspaces when the initial core basis is suboptimal or becomes unsuitable during continual learning.

Background

SPARCL freezes high-energy core directions after partitioning the accumulated autocorrelation matrix and confines subsequent adaptation to the residual subspace. The paper notes that this strategy can be harmful when the base session is small or atypical, because early core directions may be suboptimal and overly aggressive freezing may lock the learner into a poor basis. It lists delayed partitioning, trust-region-constrained core refreshes, and multiple cores for semantically distant clusters as possible practical remedies, but leaves the broader problem of adaptive repartitioning unresolved.