Can age-tuned neural networks simulate how infant senses develop?

A filter for availability, a network for extraction

The paper asks whether small neural networks tuned to a given age can turn scattered infant sensory measurements into a quick, inspectable simulation. Its answer rests on a deliberate split. What is available to a sense at a given age is set by a fixed front end: an age-indexed filter that degrades adult-quality input by the amounts reported for that age. What is learned from that input is handled by a small prediction network trained behind the filter.

This separation matters because infant data entangle three questions: what information reaches a sense, what is extracted from it, and how the timing of experience changes what is learned. A filter alone can answer only the first. A network alone has no controlled input. Placing one behind the other lets the authors compare what the input allows with what the network actually extracts at each age.

The scope is deliberately narrow. The paper models sensory perception, not cognition: object permanence, goals, number and word learning are outside it. Where the models touch on such questions, the authors say so. All data are synthetic, and the main networks cover the first 12 months, with a follow-up run to 24 months.

Each sense starts at a different point

The synthesis places the best documented sensory milestones of the first two years on one axis. The order is touch and the chemical senses first, hearing next, vision last. Touch is well developed at birth, when newborns respond to touch and show rooting and grasp reflexes, so it is the sense least changed by the simulation and is modelled with only mild blur and noise. Taste and smell also work before birth but are not modelled at all, a gap the authors state on every page that discusses senses.

Hearing begins in the womb, filtered by the mother's body, so rhythm and melody pass better than fine phonetic detail. Newborns still need sounds much louder than adults do: at 1 month, about 35 dB more at 4 kHz. Many basic abilities, including frequency resolution, are roughly adult-like by about 6 months, while sound localisation and listening in noise take longer.

Vision is the least mature sense at birth and changes most in the first year. Newborn acuity is roughly 1 cycle per degree, about 20/600 in Snellen terms, and contrast sensitivity is low. Two measurement families disagree on how fast acuity then improves: behavioural preferential-looking methods give lower values than visual evoked potentials at every age, with a much steeper rise in the first year for evoked potentials. The paper keeps both tables rather than choosing one.

Vision extracts far less than its input allows

The most decision-relevant comparison in the paper is between what a sense's filtered input contains and what the trained network can read out of it. The authors measure this with a linear probe, a ridge regression that reads the toy's position or identity from the network's state, and compare it with an information ceiling: the held-out score of a network of the same encoder family trained directly, with supervision, to recover the toy's position and identity from one filtered frame.

In vision the gap is large at early stages. At the newborn stage a toy's position cannot be read out from the network's state, with R² = 0.00, against 0.57 for a network trained directly on the same input. The network is therefore not merely limited by blur; it fails to extract information that the filtered input still contains. The authors interpret this as a property of learning and suggest that early visual limitations in infants may also be partly limitations of extraction, a hypothesis they say can be tested by comparing behavioural and evoked-potential acuity with the same stimuli.

In hearing the pattern differs: the learned read-out tracks what its input allows at every stage. The same architecture and training recipe thus produce a vision-specific gap and a hearing-specific match, which is why the paper reports per-sense comparisons rather than one aggregate score.

Critical periods emerge without being switched on

The networks also produce history effects that were not built into them. In one continuous run of a plain prediction network, a critical period and perceptual narrowing appear without anything being switched on or off at a particular age. Narrowing is measured in a run where only two of four identities ever appear, tracking surprise at identity swaps and the separability d′ of the two identities' representations during training.

The authors rank these as the most interesting computational results, precisely because the architecture contains no age-triggered mechanism. The maturation order, by contrast, is reproduced by construction: it is written into the front-end tables, so it is not evidence that networks develop like infants. The authors note that a cruder measure based on available information does not recover that order at all.

Keeping these three results apart is essential. The order of maturation is an input assumption; the slow early rise of vision beyond what the input allows is a learning effect; critical periods and narrowing are history effects that emerge from training. Only the latter two speak to how experience shapes extraction.

Hearing anchors fit early infancy but fade too fast

The authors check the hearing front end against published infant-minus-adult pure-tone thresholds. At 1 month infants are about 35 dB behind adults at 4 kHz; at 3 months about 23 dB at 500 Hz and 19 dB at 4 kHz; at 6 and 12 months about 10–15 dB, with the larger gap at low frequencies.

An exponential fit to the mean of low and high frequencies, 36.1e^(−a/2.10) + 12.6 dB for age a in months, agrees with the anchors within 3 dB at 3 and 6 months but is 12 dB higher at 1 month and 7 dB higher at 12 months. The data therefore support a steeper fall over the first months and a residual gap of about 13 dB persisting through the first year, where the anchors fall towards adult values. The data also do not support a large high-frequency cut: at 3 months the deficit at 4 kHz is smaller than at 500 Hz.

The trained hearing networks used the original anchors and were not re-run with the steeper fit. Because their learned read-out tracked the ceiling at every stage, the authors expect a stronger early attenuation to lower both together rather than open a gap, but they present this as an expectation, not a result. Touch has no comparable threshold series across the first year in the sources the authors could verify.

A second year of training adds little in this world

The follow-up continues the same full-size vision and hearing networks from 12 to 18 and 24 months, with five seeds each. Vision was run with two acuity tables: the evoked-potential table, constant after 12 months, and the website's behavioural table, which keeps rising to 10.5 cycles per degree at 24 months.

With the evoked-potential table, position read-out barely changed: 0.85 ± 0.04 at 12 months, 0.85 ± 0.05 at 18 months and 0.88 ± 0.02 at 24 months. With the behavioural table it was 0.82 ± 0.06, 0.83 ± 0.08 and 0.86 ± 0.05 at 12, 18 and 24 months. Hearing stayed at 0.45 ± 0.01, close to its ceiling of 0.49. The behavioural table's continuing acuity improvement added no information the network could use, because the toy is large compared with the remaining blur; its ceiling was already 0.94 at 3 months.

This is a result about one synthetic world, not a claim that the second year is unimportant. The authors state that the second year rests mainly on the literature synthesis and the website's tables, since the networks were trained beyond 12 months only in this follow-up.

The website shows the filter, not the baby's experience

The same front ends drive the Baby Senses Lab, a free browser-only website that shows a user's own images, videos and sounds as a baby of a chosen age might sense them. Images and videos are filtered with CSS blur, contrast and saturate computed from the vision equations; sound passes through a mid/side mix, a high-shelf filter at 3 kHz, overall attenuation and a noise floor. Because acuity is an angle, how much a picture is blurred depends on viewing distance, which is why the site asks for it and defaults to the 20–30 cm at which young infants are typically viewed.

The site also offers an age profile, an activity library of 27 everyday activities with the abilities each needs, a word checker for about 100 early words, and five short observation sessions adapted from infant research methods. Sessions are stored only in the browser and can be downloaded as CSV or JSON.

The authors are explicit about what it is not. It does not screen or diagnose, and every page says so. Displays are not calibrated instruments, one session with one baby is a noisy sample, and the age tables summarise populations rather than individuals. The views simulate the limits of the sense organs and early sensory processing, not a baby's experience: babies move their eyes differently, focus poorly at some distances in the first weeks, and their brains fill in a great deal.

Trust the per-sense gap, test the rest

The paper's most defensible contribution is the per-sense comparison of available and learned information. In vision, the newborn-stage gap between R² = 0.00 from the network's state and 0.57 for a network trained directly on the same input is a concrete, inspectable result: the simulation is not just a blur filter, and early visual limits may involve extraction as well as input quality. In hearing, the learned read-out tracks its ceiling, so the same claim does not transfer across senses.

The maturation order should not be read as evidence of development-like learning, because it is written into the front-end tables. The critical period and narrowing are stronger computational findings, but they come from one continuous run in a synthetic world, and the authors present them as history effects to be examined rather than as validated matches to infant behaviour.

The authors' own next steps define the priority: train the age-tuned models on head-camera recordings from children and, for touch and body position, in an embodied infant simulator, then compare read-outs with infant behaviour using the same paradigms. Until then, the right use of this work is to inspect what reaches an age-tuned model and what it extracts, sense by sense, rather than to treat the website or the networks as a measurement of any individual baby.

Sources used in this answer