Modern vision models are usually trained as if they were born with perfect eyesight: full resolution, full colour and high contrast from the first image. Human vision develops in the opposite direction. Newborn visual acuity, contrast sensitivity and colour discrimination mature gradually. Researchers converted that developmental path into a training curriculum and found that the order of visual experience changed what the network learned to notice.[1]
The developmental visual diet is an engineered preprocessing schedule based on psychophysical evidence. It changed the features the models learned without increasing their size. The improvement came with a trade-off: stronger shape reliance and robustness could reduce accuracy on ordinary clean images.
Selected results from the developmental visual diet
| Measure | Value | Context |
|---|---|---|
| Developmental dimensions | Acuity, contrast sensitivity and colour | Training mapped a trajectory from newborn vision to 25 years of human development.Source 1 |
| Baseline shape bias | 0.34 | Full-ecoset ResNet-50 control relied mostly on texture when shape and texture conflicted.Source 1 |
| Balanced DVD model | 0.83 shape bias | Close to the human range with a 4.3 percentage-point accuracy decline from baseline.Source 1 |
| Performance-prioritised DVD model | 65.03% accuracy and 0.70 shape bias | Baseline accuracy was 62.99% with 0.34 shape bias.Source 1 |
| Abstract-shape recall | 36.21% | DVD-S on IllusionBench versus 8.71% for the ResNet-50 baseline.Source 1 |
| PGD adversarial accuracy | 39% versus 11% | Reported at the paper’s severity-five condition.Source 1 |
| High-severity corruptions | Two to four times baseline accuracy | Range reported across noise, blur, weather and image-quality deficits.Source 1 |
The curriculum changes the feature economy
When fine texture is available from the beginning, texture can become the cheapest predictive shortcut. Early low acuity and developing contrast sensitivity preserve broad structure while limiting local detail. The model is pushed to integrate information across the object before precise texture becomes reliable.
Controlled rearing experiments identified contrast-sensitivity development as the largest single contributor. Acuity-only schedules helped less, and the full combination produced the most consistent result. Chronological ordering also outperformed a shuffled schedule using the same kinds of visual degradation, suggesting that progression—not only augmentation—matters.[1]
What the result changes for vision-system design
| Design assumption | Standard practice | Developmental-diet alternative |
|---|---|---|
| Input quality | Show the cleanest available image from the first epoch | Begin with controlled limits and increase fidelity over trainingSource 1 |
| Robustness strategy | Add corruptions or adversarial training after shortcut features emerge | Shape the early features before high-frequency detail dominatesSource 1 |
| Evaluation | Measure clean-image top-1 accuracy | Add shape bias, abstract-shape recall, corruption curves and attack robustnessSource 1 |
| Scale hypothesis | More data and parameters should solve misalignment | Training order can outperform much larger systems on selected robustness tasksSource 1 |
| Deployment claim | Human-like score implies human-like perception | Inspect which cues and failure modes the model actually uses |
A small clean-image loss can buy a different failure profile
The balanced model sacrificed several percentage points of clean-image accuracy while moving much closer to human shape bias. The performance-prioritised version slightly improved clean accuracy and still more than doubled shape bias. This makes the choice an operating point rather than a single champion: a safety-critical application may value graceful degradation differently from a photo benchmark.
The study also compared developmental training with adversarial training. Adversarial training was strong against attacks similar to those it had seen, while the developmental models generalised more broadly across black-box perturbations and natural corruptions. That does not remove the need for threat-specific defences; it suggests the underlying representation can make several defences easier.
The next test is task realism
Industrial vision systems face motion, lighting, sensor changes and rare objects, not only benchmark corruptions. A serious pilot would train matched systems, then test camera shifts, weather, compression, deliberate attacks and domain-specific mistakes. Reviewers should ask whether the new failure pattern is safer—not merely whether the average score rose.
The developmental visual diet is compelling because it converts a fact about human maturation into an experimentally testable design choice. It offers a different route to robust AI: do not only repair shortcuts after they form. Control the sequence of experience that makes the shortcuts attractive in the first place.
Sources & further reading
- Adopting a human developmental visual diet yields robust and shape-based AI vision Nature Machine Intelligence · 24 Apr 2026 · Accessed 11 Oct 2026