When is an old example forgotten? Predicting first-misclassification times from the tangent kernel at the task switch

Preprint, under consideration at Pattern Recognition Letters. We evaluate the empirical neural tangent kernel frozen at the task switch as a predictor of which old examples a two-layer ReLU network first misclassifies under the cross-entropy loss, and when, with random labels and a domain-incremental split of a handwritten-digit dataset. The archive contains the manuscript, the Supplementary material, all code, per-seed results, the protocol fixed before the fresh seeds were run (with a dated explanatory note), and scripts that reproduce every number and figure (see README.txt).