Deep neural networks have critical learning periods like Hubel–Wiesel's kittens — an early input deficit becomes permanent

1. Deep nets have critical periods, timed like animal ones. A temporary corruption of a net's inputs early in training can permanently cap the final skill, and the damage scales with when and how long the deficit lasts — not the total exposure.

"Similar to humans and animals, 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" — arXiv:1711.08856 (Tier 1).

2. The biological anchor is a "fix-it-late-and-you're-too-late" result. Hubel & Wiesel sutured one kitten eye shut; the open eye's cortical columns permanently annexed the deprived eye's territory even though the eye itself was healthy. "Ocular dominance is established irreversibly early in childhood" — PMC11445666 (peer-reviewed review, Tier 1/2). This is why pediatric surgeons now remove congenital cataracts within weeks: the clock is in the cortex, not the lens.

3. The mechanism is which information, when — not how much. Deficits that spare low-level statistics are recoverable; blur (a cataract analogue) is not. Achille et al. tie this to a rise-then-fall of Fisher Information: "Information rises rapidly in the early phases of training, and then decreases … a phenomenon we refer to as a loss of 'Information Plasticity'." (Tier 1). Crucially, the net has no synaptic pruning or neuromodulators — the critical period is a property of learning dynamics, not biological hardware.

Why this was hop-worthy

A 1960s cat-vision accident (and a Nobel) turns out to describe how a 2019 deep net fails — a cross-domain, cross-time bridge that lands right back on AI training dynamics.

Further leads

• Achille's "Information Plasticity" is the Information-Bottleneck view (Tishby: "the most important part of learning is actually forgetting") applied to critical periods — does the vault link IB to critical periods?
• Deep-net critical period ↔ the vanishing gradient problem (both: early layers stop being able to change).
• Loop-closure candidate: has the DL model ever predicted human amblyopia critical-period timing (biology → DL → back to clinic)?

Hop chain

Chain: Neocognitron → deep-net critical learning periods

Hop 1: "An accidental experiment discovered new cells in cat brains…" — https://massivesci.com/notes/simple-complex-cells-neurons-cats-eyes/

• Hook type: Cross-domain bridge (neuroscience → the seed's CNN architecture)
• Hook: The Neocognitron's S-cells/C-cells descend from Hubel & Wiesel's simple/complex cortical cells — discovered by accident when a neuron fired at the moving edge of a glass slide, not the projected dot.
• Why followed: Highest-priority hook type; it leaves the CNN topic into the neuroscience it was copied from.
• Key findings: 1959, Johns Hopkins; simple cells fire for lines at a specific orientation, complex cells for oriented lines moving in a direction. Mentor Stephen Kuffler (center-surround receptive fields) set it up.

Hop 2: "From Cats to the Cortex" — https://pmc.ncbi.nlm.nih.gov/articles/PMC11445666/

• Hook type: Surprising claim + cross-domain bridge (neuroscience → clinical medicine)
• Hook: Depriving one healthy eye during a critical window causes permanent deficit — the loss is in the cortex, not the eye.
• Why followed: A reversal of the naive "the eye is fine so vision recovers" prior, with a real clinical payoff.
• Key findings: Ocular dominance fixed irreversibly early; critical period identified mid-1960s; revolutionized management of congenital cataracts and amblyopia (treat early or not at all).

Hop 3: "Critical Learning Periods in Deep Networks" — https://arxiv.org/abs/1711.08856

• Hook type: Cross-domain + cross-time bridge (the road home to AI)
• Hook: Deep nets show the same onset/length-dependent critical periods as animals.
• Why followed: Bridges the 1960s cat result directly to modern training dynamics; near the vault's vanishing-gradient and delayed-vindication notes.
• Key findings: Blur deficit (cataract analogue) early in training is unrecoverable; vertical-flip deficit is recoverable; explained by loss of "Information Plasticity"; no biological machinery needed.

Hop 4: "New Theory Cracks Open the Black Box of Deep Learning" — https://www.quantamagazine.org/new-theory-cracks-open-the-black-box-of-deep-learning-20170921/

• Hook type: Mechanism question (what is "Information Plasticity," really?)
• Hook: Tishby's Information Bottleneck — training's long second phase compresses/forgets input detail to generalize.
• Why followed: Achille & Soatto are IB theorists; this grounds the "information rises then falls" mechanism.
• Key findings: Two phases (fit then compress); "the most important part of learning is actually forgetting." But Andrew Saxe showed some large nets generalize well without a drawn-out compression phase — the phase may not be universal.

Saved hooks not followed:

• Stephen Kuffler — Hubel & Wiesel's mentor, discovered retinal center-surround fields, "father of modern neuroscience," never won a Nobel — from Hop 1 — a strong "person behind the thing" hook, deferred to keep the chain on plasticity.
• Golden Goose Award (obscure federally-funded cat research → major medical advances) — from the Hubel/Wiesel search — an institutional/cultural hook about how basic science pays off unpredictably.
• Lottery-ticket / early-phase-of-training determinism (novelty 0.724) — a sibling "the first epochs decide everything" thread, different mechanism (trainable subnetworks), saved as its own future chain.

Surprise: expected the deprived eye itself to be damaged — found the eye is healthy and the permanent deficit lives entirely in cortical wiring. Surprise: expected biological critical periods to need molecular machinery (pruning, neuromodulators) — found a plain deep net with none of that reproduces the same onset/length-dependent critical period. Surprise: expected "more training always helps recover" — found blur-deficit damage is permanent while a vertical-flip deficit fully recovers; deficit type, not just duration, decides. Surprise: expected compression-to-generalize to be the settled story of deep learning — found Saxe's result that some nets skip the compression phase and still generalize.

post-worthy: maybe — a clean 1960s-cats-to-2019-nets bridge with a genuine deflationary twist (no molecules required), but the human-critical-period ↔ deep-net link deserves one more primary source before publishing.