Deep neural networks exhibit critical learning periods — a temporary input deficit early in training can permanently cap the final skill, timed by onset and length like animal critical periods

A temporary corruption of a deep network's inputs early in training can permanently cap its final skill, and the size of the damage scales with when the deficit begins and how long it lasts — not with the total amount of degraded exposure. Achille, Rovere & Soatto (2019) show this directly: "deep artificial neural networks exhibit critical periods during which a temporary stimulus deficit can impair the development of a skill … The extent of the impairment depends on the onset and length of the deficit window, as in animal models." Restore clean inputs after the window has closed and the network never fully recovers, even with unlimited further training.

This mirrors, in an artificial system, the developmental critical periods of biology — most sharply the monocular-deprivation result that a healthy but deprived eye loses cortical territory forever if the deprivation falls inside an early window (claim-monocular-deprivation-permanently-rewires-visual-cortex). The two systems share the same onset/length-dependent signature, which is why the finding reads as a cross-domain, cross-time bridge: a 1960s cat-vision result describing how a 2019 net fails.

The claim as stated here is validated against animal critical periods; whether a deep net's critical-period timing has ever matched a human one (e.g. amblyopia) is a separate, open loop-closure question (question-deep-net-critical-period-predicts-human-amblyopia-timing). The mechanism behind the phenomenon — a rise-then-fall of information rather than any biological machinery — is treated separately in claim-critical-periods-arise-from-information-plasticity-not-biology. The "early layers lose the ability to change" character also rhymes with the vanishing-gradient pathology (claim-vanishing-gradient-chain-rule-pathology), though the two are distinct failure families. The CNN substrate these experiments run on descends architecturally from the same Hubel–Wiesel neuroscience the biological anchor comes from (claim-fukushima-1979-neocognitron-first-cnn).