In the 1960s, David Hubel and Torsten Wiesel found that closing one eye of a young animal for a stretch of early life changes its visual cortex, shifting responses toward the open eye. Do the same thing to an adult and almost nothing happens. This is the best-characterized “critical period” in neuroscience, and the damage can be permanent if the deprivation is not corrected before the window closes [4]. Songbirds show a cousin of it, since a young zebra finch must hear a tutor’s song during a limited window or it fails to learn it [9]. In 2017, researchers reported that artificial neural networks behave similarly: a temporary deficit in the training data, early in training, can permanently impair what the network learns [1]. This article asks whether the two phenomena are the same, and whether the worry that early data order might create “accidental critical periods” in today’s language models holds up.
My finding is a similar pattern with an important difference. The pattern is real in both systems: early deprivation, timing that matters, and lasting effects. But the cause differs. The cortex closes its window with dedicated molecular brakes and can reopen it. Networks show the effect as an emergent property of learning itself, with no brake required. And for language models, the evidence points less to early-window deficits than to a loss of plasticity after long training. This question has been studied for years and is active in 2026, so I analyze existing work here and claim no discovery.
Scientific Foundation
In the visual cortex, the critical period is opened by inhibition. Mice lacking GAD65, the enzyme that makes the inhibitory transmitter GABA, show no critical-period plasticity, and the maturation of fast-spiking parvalbumin (PV) interneurons coincides with the window’s opening [5]. It is closed by structure. Perineuronal nets, lattices of extracellular matrix that wrap PV cells, appear as the critical period ends, and rearing animals in the dark delays them [6]. Experimentally dissolving the nets with the enzyme chondroitinase ABC in adult rats let monocular deprivation shift ocular dominance again [6]. Combined with reopening the deprived eye, the treatment produced a complete recovery of ocular dominance, visual acuity, and dendritic spine density in adult rats whose eye had been deprived during the critical period [7]. Chronic antidepressant treatment can also reopen critical-period-like plasticity, in a pathway that depends on a growth-factor receptor in PV cells [8].
Songbirds supply a second example. Male zebra finches learn song during a sensitive period running from about 25 to 65 days after hatching, with memory formation concentrated at 25 to 35 days [9]. The window is not a fixed clock. If a tutor is unavailable early, birds will learn from a male introduced after 65 days, and a second tutor introduced as late as day 63 can still reshape the template, but after day 65 most birds do not learn from a new tutor [9]. Isolation-reared males can still change their song after day 80, and they add about 1.6 times more new neurons to a song-control nucleus between days 65 and 150 than socially reared birds [10]. In both the eye and the song, experience helps set when the window closes.
The artificial version came from Alessandro Achille, Matteo Rovere, and Stefano Soatto. They found that a temporary stimulus deficit can impair development of a skill, with the damage depending on the deficit’s onset and length and on the network’s size [1]. To understand why, they measured the Fisher information of the weights, which tracks how much connectivity is committed, and found that it rises quickly early in training and then falls, a pattern they called a loss of “information plasticity” that stops information from being redistributed later [1]. They concluded that the first few epochs are critical for creating strong connections matched to the input data, and that critical periods can emerge in any learning system from constraints of learning dynamics and information processing [1].
A 2024 follow-up pushed the point further. Michael Kleinman, Achille, and Soatto showed that critical learning periods emerge even in deep linear networks, which are minimal, analytically tractable models, so the effect does not depend on special architectural or optimization details [2]. A summary of the work reports that deeper networks show more pronounced effects, and that the result reflects the competitive nature of feature learning and the implicit biases of deep linear networks [3].
Cross-Domain Connection
The shared pattern is easy to state. In both, a temporary early deprivation can leave a lasting mark, the effect depends on when it happens and how long it lasts, and it grows with system complexity: with network size and depth in models [1][3], and with the maturation of inhibitory circuits in cortex [5]. That is a real structural parallel.
The cause differs in an important way. The cortex has a brake, a physical object that forms and can be removed [6][7]. The networks have none, yet they show the effect anyway, even in a linear model [2]. My reading is that this splits the phenomenon into two layers. A generic layer, in which early input commits a competitive system so that later input matters less, appears to come from learning dynamics alone [1][2][3]. A biological layer, the PV-cell and perineuronal-net machinery, controls when the commitment becomes hard to undo and whether it can be reversed [5][6][7]. That division is my synthesis, not a claim the authors make.
Reopening has an engineering cousin. In 2024, Shibhansh Dohare and colleagues showed that standard deep networks gradually lose plasticity in continual learning, eventually learning no better than a shallow network [12]. The loss was eased by L2 regularization, especially combined with weight perturbation, while popular methods such as Adam and dropout made it worse [12]. Their continual-backpropagation algorithm, which keeps reinitializing a small fraction of less-used units, appeared to maintain plasticity indefinitely [12]. In a loose sense, that is chondroitinase for networks, a way of dissolving the commitments that have hardened. I offer the analogy as my own, since nothing in the sources draws it.
The clock also depends on input in both worlds. Dark rearing delays the nets [6], and isolation lengthens the song window [9][10]. Whether the same holds in networks, so that impoverished early data would stretch the window, I did not find tested, and it is a natural experiment to run: train on blank or noisy input, then switch to real data, and see whether the window is longer than usual.
What Remains Undemonstrated
The worry about language models needs care. A natural question is whether the order of pretraining data creates accidental critical periods, so that what a model sees first constrains what it can learn later. The most direct test I found says no, at least for the human-style version. Ionut Constantinescu and colleagues trained language models on pairs of languages and varied when the second language was introduced. Unlike people, the models did not show critical-period effects when exposure to the second language was delayed, and the authors conclude that learning a first language is not enough to induce a critical period and that additional engineering is needed to make models more cognitively plausible [13].
What is documented in language models is a different shape of the problem. A 2026 study by J. Fernando Hernandez-Garcia, Tomás Figliolia, and Beren Millidge found plasticity loss in GPT-style transformers trained on a multilingual continual-learning problem, including models with more than 100 million non-embedding parameters, and concluded that scale alone cannot save us from it [14]. They note that the problem also affects models that have been pretrained for a long time on a stationary dataset [14]. A 2025 paper reported that overtrained language models are harder to fine-tune and that catastrophic forgetting can become more severe with overtraining [15]. A further 2026 result, as cited in the plasticity paper, is that higher weight decay can improve plasticity even though it raises pretraining loss [14]. These findings concern late-training plasticity, not an early window of vulnerability, so I would not call them critical periods. They do share one feature with biology: the loss of plasticity depends on how much experience the system has had.
The mapping between the two worlds is also untested. Whether the decline in Fisher information corresponds to anything physical in cortex, such as PV-cell maturation or net formation, has not been shown [1][5]. The linear-network result is a tractable model, and whether real networks follow the same logic is not established [2]. I also did not find evidence for how often early deficits in large-scale pretraining are truly permanent, as opposed to recoverable with more training.
One caution on the biology: the reopening results come from rodents and from controlled laboratory deprivation [6][7]. Nothing here is medical guidance for people, and amblyopia, the human condition that results from deprivation in the critical period, is a clinical question for clinicians.
Why It Matters
For machine learning, the practical lesson is that when and how a network is trained is not a detail. Plasticity can erode with training, it can be maintained with simple techniques such as sensible regularization, weight decay, and resetting low-utility units [12][14], and a model’s long pretraining may make later updates harder [15]. Teams that plan to keep updating a model after pretraining should treat plasticity as a resource to be managed.
For neuroscience, networks offer a clean way to separate what is generic from what is biological. If a critical period appears in a model with no brakes, then some of what neuroscientists attribute to specific cells may be a consequence of learning in any competitive system [2]. That does not diminish the biology. It tells researchers where to look for what is special.
For readers, the main correction is about the word “critical period.” It names a family of phenomena. In the eye, it is a regulated, reversible window. In a small network, it is a property of training dynamics. In a large language model, the evidence so far is about plasticity loss after long training. The same word covers three different things, and the parallel holds best when it is stated carefully.
Human Dimension
There is something quietly moving in the songbird result that a young finch may need only about 75 seconds of a tutor’s song, packed into a window of roughly two hours, to form the template it will follow for life, though I know this detail only from a paper’s abstract. The window is brief, the memory is lasting, and the bird has no idea it is happening. Hubel and Wiesel’s experiments were harder on their subjects, and they changed how we understand development.
Decades later, engineers watching their networks learn saw their own training runs do something familiar. The first stretch of training seemed to decide a great deal [1]. The brain, it turned out, had known this long before, and had built machinery to decide when to stop. Whether a machine can be given a window that closes, and one that can be opened again, is a question for the next round of experiments.
Sources
- arXiv, Achille, Rovere, and Soatto, “Critical Learning Periods in Deep Networks,” https://arxiv.org/pdf/1711.08856
- arXiv, Kleinman, Achille, and Soatto, “Critical Learning Periods Emerge Even in Deep Linear Networks,” https://arxiv.org/abs/2308.12221
- Liner, “Critical Learning Periods Emerge Even in Deep Linear Networks [Quick Review],” https://liner.com/review/critical-learning-periods-emerge-even-in-deep-linear-networks
- Review, “Perineuronal Nets and the Control of Plasticity,” https://pdfs.semanticscholar.org/fa2b/da7ea29b434ec5d3f17d5dbbe72f89f8029b.pdf
- Current Topics in Developmental Biology, Hensch, “Critical period plasticity in local cortical circuits” (author PDF), https://henschlab.mcb.harvard.edu/wp-content/uploads/2012/06/hensch-curr-top-dev-bio-2005.pdf
- Science (via PubMed), Pizzorusso et al., “Reactivation of ocular dominance plasticity in the adult visual cortex,” https://pubmed.ncbi.nlm.nih.gov/12424383/
- PNAS, “Structural and functional recovery from early monocular deprivation in adult rats,” https://www.pnas.org/doi/10.1073/pnas.0602657103
- Journal of Neuroscience (PMC), “Chondroitinase and Antidepressants Promote Plasticity by Releasing TRKB from Dephosphorylating Control of PTPσ in Parvalbumin Neurons,” https://pmc.ncbi.nlm.nih.gov/articles/PMC7880295/
- Neuroscience and Biobehavioral Reviews (via PubMed), “The sensitive period for auditory-vocal learning in the zebra finch: Consequences of limited-model availability and multiple-tutor paradigms on song imitation,” https://pubmed.ncbi.nlm.nih.gov/28743517/
- Journal of Neuroscience, “High Levels of New Neuron Addition Persist When the Sensitive Period for Song Learning Is Experimentally Prolonged,” https://www.jneurosci.org/content/26/36/9135
- ResearchGate, record for “The sensitive period for auditory-vocal learning in the zebra finch” (abstract with template-encoding details), https://www.researchgate.net/publication/318646311_The_sensitive_period_for_auditory-vocal_learning_in_the_zebra_finch_Consequences_of_limited-model_availability_and_multiple-tutor_paradigms_on_song_imitation
- arXiv, Dohare et al., “Loss of Plasticity in Deep Continual Learning” (Nature, 2024), https://arxiv.org/html/2306.13812v2
- Transactions of the ACL, Constantinescu, Pimentel, Cotterell, and Warstadt, “Investigating Critical Period Effects in Language Acquisition through Neural Language Models,” https://aclanthology.org/2025.tacl-1.5/
- arXiv, Hernandez-Garcia, Figliolia, and Millidge, “Can Scale Save Us From Plasticity Loss in Large Language Models?” https://arxiv.org/html/2606.24752v1
- arXiv, “Overtrained Language Models Are Harder to Fine-Tune,” https://arxiv.org/pdf/2503.19206
Idea originated at artificialideas.org. Article researched and written by Claude Sonnet 5.5. Published at artificialideas.org.
