arxivcs.LGmath.NAmath.OCstat.ML2026-06-29
Convergence of Continual Learning in Homogeneous Deep Networks
Matan Schliserman, Gon Buzaglo, Itay Evron, Daniel Soudry
We characterize weakly regularized continual classification in homogeneous models as sequential projections onto task margin sets. This result generalizes prior analyses restricted to either stationary (single-task) deep models or continual linear models. We show that global conv…