Analytic continual learning (ACL) methods such as ACIL and GACL fit a ridge-regression classifier head on frozen features. They use the Sherman-Morrison-Woodbury identity to produce exemplar-free, exact updates, avoiding the need to store previous data.
The paper focuses on the use of equiangular tight frames (ETFs) as the target geometry for the classifier head. It reports an impossibility result for that geometry in the ACL setting, indicating a fundamental limitation rather than a mere implementation issue.
As a remedy, the paper proposes an equinorm adjustment for imbalanced analytic continual learning. This is meant to keep class target norms equal when class frequencies are uneven. The abstract does not include experimental details, but the theoretical result is clearly stated.