Findings

What we found

The main results of both papers, each stated with its numbers. Values are from the current versions of the papers.

Scope. All results come from small networks in a simple synthetic world with one toy, scored mostly with linear read-outs. They describe this setting; whether they hold at scale or with real video is not tested. Scores are R² (the share of a quantity's variance a linear read-out recovers from the model; 1 is perfect, 0 is no better than the average), given as mean ± s.d. over independent training seeds.

Newborn-quality input slows vision far below what the input allows

At the first stage the vision model fed newborn-quality input cannot locate the toy, although a network trained directly on the same input can. It catches up as the input sharpens.

Toy position at the first stage: 0.02 ± 0.00, against 0.60 available in the input. At the 12-month stage: 0.85 ± 0.04, against 0.89 ± 0.03 with adult-quality input throughout. 5 seeds

Contrast, not blur, produces the early lag

For a linear read-out we prove that scaling contrast by κ delays learning by κ⁻², while blur lowers the reachable ceiling. Front-end ablations in the trained network agree.

First-stage position: blur only 0.34 ± 0.05; contrast only 0.04 ± 0.02; full schedule 0.02 ± 0.00. Colour was held at adult values in both ablations, so its role is untested. 5 seeds

A gentle start helps hearing a little and costs vision a little

Models that heard degraded sound early localise slightly better at the end. Shuffled input quality helps about as much, so the gain looks more like augmentation than an effect of order.

Hearing localisation at 12 months: 0.45 ± 0.01 with immature input, 0.39 ± 0.05 with adult input (g = 1.5, p = 0.05). Vision: 0.85 ± 0.04 against 0.89 ± 0.03 (p = 0.06). 5 seeds

Growing the network halves compute but slows learning

Widening the networks between stages, without changing what they compute, delays learning after each widening, as the growth proposition predicts.

Compute: 0.48 of a fixed-size network. Vision at 12 months: position 0.79 ± 0.06 grown against 0.85 ± 0.04 fixed; identity 0.49 ± 0.06 against 0.64 ± 0.06 (g = −2.2, p = 0.005). 5 seeds

Combining senses tracks a hidden toy best

While the toy is behind a screen, its position is best decoded from the fused world model. The advantage appears from the 6-month stage.

Hidden toy position at 12 months: fused 0.75 ± 0.06, vision memory 0.60 ± 0.08, hearing 0.42 ± 0.09; fused beat vision in every seed. 5 seeds

Decodable is not expected: no vision model is surprised when a hidden toy vanishes

Every unimodal vision model is more surprised when the toy reappears than when it fails to, although its position can be read from the model's memory. The fused model is surprised only when the toy's sound stops. A forecaster that also predicts its own uncertainty does not change this.

Fused surprise index 0.32 ± 0.11 (hearing channel 0.75 ± 0.05, vision channel −0.04 ± 0.05); silent toy −0.04 ± 0.02. Uncertainty-aware forecaster −0.35 ± 0.23. A latent forecaster did not learn the toy, so its test is uninformative. 5 seeds

Matching experience to age leaves the 12-month result unchanged

In the main schedule each stage gets the same amount of training. Making experience proportional to age changes the early stages but not the end point.

Vision at 12 months: 0.86 ± 0.01 matched against 0.85 ± 0.04 staged; hearing 0.45 ± 0.01 in both. 5 seeds

From Learning Like Infants: Input Constraints in Predictive World Models. Statistics: Welch's t-test and Hedges' g at the 12-month stage.

The filters alone put the senses in the infant order, by construction

Touch, then hearing, then vision appears without any change to the network, because the order is in the age tables. Measured instead by when the available information reaches half its final value, the infant order is not recovered.

Half-ceiling ages: touch 0, hearing 4, vision 0 months; Kendall τ = 0.00 against touch before hearing before vision. information ceiling

Vision networks learn less than their input allows; hearing networks about as much

Separating the age filter from the learner shows that early vision is limited twice, by the input and by how slowly a generic prediction objective picks out a small, faint toy.

Vision position, learned against available: 0.00 ± 0.01 vs 0.57 (newborn stage), 0.29 ± 0.16 vs 0.93 (3 months), 0.85 ± 0.04 vs 0.95 (12 months). Hearing at 12 months: 0.45 ± 0.01 vs 0.48. 5 seeds

A critical period and perceptual narrowing emerge from plain prediction

A window of heavy blur leaves a lasting loss of fine detail only when it comes first. A network that only ever sees two of four identities first gains, then loses, sensitivity to the unseen pair.

Fine-detail accuracy after a 3-month blur window at the start: 0.81 ± 0.04, against 0.94 to 1.00 for later windows. Surprise at the unseen pair: peak 0.54 ± 0.01, falling to 0.29 ± 0.09. 3 seeds

The first sense to come online can crowd out a later one

When two senses share one read-out and hearing starts first, as in infants, the vision encoder ends up under-trained. Training each sense separately and combining them later gives the best encoders.

Vision encoder with hearing first: 0.60, against 0.84 for vision alone; separate training then combining: 0.84 and 0.84. companion study

A second year of training adds little in this world

Continuing the networks to 24 months barely changes toy position, which is already near its ceiling, while toy identity in vision keeps improving with experience.

Vision position 0.85 ± 0.04 at 12 months, 0.88 ± 0.02 at 24; identity 0.64 ± 0.06 to 0.70 ± 0.09. Hearing stays at 0.45 ± 0.01 (ceiling 0.49). 5 seeds

Latent prediction did not find the toy in this small setting

Predicting the next latent code instead of the next frame collapsed without regularisation and, with variance and covariance penalties, kept a spread-out code that carried almost nothing about the toy.

Vision position at 12 months: 0.02 ± 0.04 latent prediction (regularised) against 0.85 ± 0.04 pixel prediction. 5 seeds

The hearing filter matches infant thresholds in mid-infancy only

Compared with published infant-minus-adult thresholds, the hearing settings agree within 3 dB at 3 and 6 months and are too mild at 1 month and at 12 months.

Fitted gap: 36.1·e^(−age/2.10) + 12.6 dB; the settings are 12 dB too mild at 1 month and 7 dB at 12 months. published data

From Simulating Infant Senses with Age-Tuned Networks.