Recovery on Signal or on Schedule? A Small Pre-Registered Test of Health-Preserving Pedagogy in Continual Learning

Anton Dziatkovskii · ORCID 0000-0001-7408-3054

Preprint v2, 9 October 2026. Version of record: Zenodo, DOI 10.5281/zenodo.23268660. License CC BY 4.0.

Full text (PDF, 7 pages, 79 KB)DOICode, seeds and raw tables: machine-pedagogy

Abstract

Health-preserving pedagogy holds that teaching has a cost to the learner and that a recovery phase should start when the learner's diagnosed state calls for one, not when the timetable does. We test the machine analogue at a fixed gradient-step budget in class- and domain-incremental continual learning (5 tasks from the 8×8 digits set, a small MLP).

A threshold arm replays a small memory only when a held-out probe of past tasks shows a loss rise above tau. We compare it with replay of exactly the same count placed five other ways: uniformly over the whole run, at random, at the start of each task, at random inside each task, and evenly inside each task with the threshold arm's own per-task counts. A sixth comparison is a sensor-free rule that makes replay proportional to the number of past tasks.

On a pre-registered fresh holdout (seeds 20–29) at equal cost, the threshold arm beats globally uniform replay. On class-incremental Split-Digits the gain is 3.4–7.7 points (8–10 of 10 seeds at tau 0.5–1.0); on Permuted-Digits it is 0.6–4.1 points and reaches 8–10 of 10 seeds only at tau ≥ 0.75. The sensor-free proportional rule is never beaten by the threshold arm at any of the ten points and exceeds it by 6.2 points at the smallest budget on Split. An oracle control that keeps the sensor's per-task counts but spaces them evenly matches the threshold arm at medium budgets and exceeds it at the two smallest.

What transfers from the pedagogy is the dosage rule: more recovery as more material accumulates, spread evenly. On this benchmark the state alarm adds nothing measurable beyond what a fixed proportional rule already gives, and at small budgets its within-task timing costs accuracy. Code, seeds, raw tables, the pre-registration with two dated addenda and two superseded versions are public.

Keywords

continual learning · experience replay · catastrophic forgetting · curriculum learning · health-preserving pedagogy · homeostatic approach · pre-registration · negative result · recovery phase · learner state · cost of learning

How to cite

Dziatkovskii, A. (2026). Recovery on Signal or on Schedule? A Small Pre-Registered Test of Health-Preserving Pedagogy in Continual Learning. Preprint v2, Zenodo. https://doi.org/10.5281/zenodo.23268660

@misc{dziatkovskii2026recoveryonsignal,
  author    = {Dziatkovskii, Anton},
  title     = {Recovery on Signal or on Schedule? A Small Pre-Registered Test of Health-Preserving Pedagogy in Continual Learning},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23268660},
  url       = {https://doi.org/10.5281/zenodo.23268660},
  note      = {Preprint v2, CC BY 4.0}
}

One text, one DOI: the PDF on this page is the Zenodo file byte for byte. Please cite the DOI.