The founding deep-network critical-period paper validated its timing against animal data only, and never attempted a human match

critical-periodsdeep-learningamblyopianeuroscienceloop-closureplasticitycross-domain-bridge

The paper that established critical periods in deep networks — Achille, Rovere & Soatto (2019) — fits its deep network's sensitivity-to-deficit curve against animal data: monocularly-deprived-kitten results (Olson & Freeman 1980; Giffin & Mitchell 1978) and macaque synaptic-density-over-age data (Rakic et al. 1986). It does not fit the curve against any human clinical dataset. Even the animal comparison is a substitution forced by data availability: the deficit the network simulates is a cataract, but the fitted curve comes from monocular deprivation instead, because "while the overall trends of cataract-induced critical periods have been studied and understood in animal models, there is not enough data to confidently regress sensibility curves comparable to those obtained in DNNs. For this reason, in Figure 1 we compare the performance loss in a DNN trained in the presence of a cataract-like deficit with the results obtained from monocularly deprived kittens" (Appendix C). Read in full, that sentence is about the sparseness of cataract-model data in animals. It is not a statement about human clinical data, and the paper makes none.

Human amblyopia enters the paper only qualitatively, as motivating background citing clinical literature (von Noorden 1981; Taylor et al. 1979) for the point that treatment outcome depends on both the duration of the deficit and its age of onset. That qualitative dependence is never turned into a curve the DNN's own sensitivity profile is fit or compared against.

This scopes the vault's headline critical-period result. The match that claim-deep-nets-have-critical-learning-periods-timed-like-animals records is an animal match — the onset/length signature the net shares with the monocular-deprivation result (claim-monocular-deprivation-permanently-rewires-visual-cortex). The mechanism the net reproduces without biological hardware is treated in claim-critical-periods-arise-from-information-plasticity-not-biology. What this note fixes is the ceiling on how far the founding paper itself carried the claim toward human timing: not at all — silently, without offering a reason. Whether any later work has closed that gap is the open loop-closure thread question-deep-net-critical-period-predicts-human-amblyopia-timing, answered provisionally in claim-no-dnn-model-has-matched-human-critical-period-timing.

Correction history.

• 2026-08-25 (cross-model audit, claude-opus-5) — Reached incidentally while auditing moc-a-critical-period-that-means-less-than-it-shows, which carried the same error and has also been corrected. The note previously read that the paper stated "human clinical data was too sparse to regress a comparable curve" (title) and that it carried the claim toward human timing "not at all, and by explicit admission of missing data" (body), with a commentary built on that reading: "it says the human curve couldn't be drawn, because the data to fit it against wasn't there… It's unclosed because the clinic hadn't measured finely enough in 2019 for the net to have anything to line up against." The paper says no such thing. Its source_quote had been truncated to begin at "there is not enough data", which severed the clause naming the referent — "cataract-induced critical periods… in animal models". Quote restored to the full sentence; title, body, and commentary corrected. The claim the note exists to make — animal-only validation, human amblyopia as qualitative motivation only — was independently re-verified against the paper and stands.