Every major AI lab has spent the last decade betting on the same idea: if you want better machine vision, train on more images with more computing power. A study published in Nature Machine Intelligence challenges that assumption directly. The researchers found that the order and quality of what an AI learns matters more than the quantity. By mimicking the way human vision develops from birth, they produced AI systems that see more like humans do, resist attacks that fool standard models, and do so without the massive computational cost of scaling up.
The implications reach into every system that depends on machine vision: autonomous vehicles, medical imaging, satellite analysis, security systems, and robotics.
What Researchers Did
Human infants do not start life seeing the world clearly. Newborn vision is blurry, low contrast, and mostly greyscale. Over the first months of life, acuity sharpens, colour perception develops, and contrast sensitivity improves. This is not a limitation. It appears to be a feature.
A team of researchers from institutions including the Free University of Berlin encoded this developmental progression into a training curriculum for AI vision systems. They called it a developmental visual diet. Instead of feeding the AI a massive dataset of sharp, full-colour images from the start, they began with blurry, low-contrast, greyscale inputs and gradually increased the visual complexity over time, following the documented stages of human infant visual maturation.
The results, reported across a benchmark of ten datasets, were consistent. Models trained this way showed the strongest reliance on shape information yet reported in the literature. Standard AI vision systems notoriously rely on texture instead of shape. Show one a picture of a cat with elephant-skin texture applied to it, and it will often classify it as an elephant. Humans do not make this mistake because we learn to see shapes first.
The developmentally trained models also showed substantially better resistance to adversarial attacks, which are small, often invisible modifications to an image designed to fool AI systems into misclassifying objects. They performed better on abstract shape recognition, on out-of-distribution images, and on depth perception tests.
Why This Matters
The dominant assumption in AI development is that scale solves problems. More data, more parameters, more compute. This paper demonstrates a specific case where that assumption fails, and where a different approach, guided by neuroscience rather than engineering instinct, produces better results.
The practical stakes are significant. AI vision systems that rely on texture rather than shape are fragile in real-world conditions. Lighting changes, weather, image compression, and adversarial interference all affect texture more than shape. A medical imaging system that identifies tumours by texture pattern rather than structural shape may fail in ways its developers never anticipated. An autonomous vehicle that misreads a stop sign because someone placed a sticker on it represents a known vulnerability traceable to this same texture bias.
The developmental training approach does not require larger datasets or more powerful hardware. It requires a different curriculum. That distinction matters economically. Training large vision models is expensive. A method that improves reliability without increasing cost changes the deployment calculus for applications where failure is not acceptable.
What Comes Next
The researchers tested their approach on standard benchmarks and found improvements across all of them. The limitation is that this work addresses visual robustness, not all aspects of machine vision performance. Speed, accuracy on highly specific tasks, and performance in very constrained domains may not all benefit equally.
The deeper question this paper raises is whether other areas of AI training are similarly missing a lesson from biology. Infant development is not arbitrary. It is the product of millions of years of optimisation. This study suggests that when human development and AI engineering diverge, it may be worth asking which one has the better answer.
Sources
Lu, Z., Thorat, S., Cichy, R.M. et al. (2026). Adopting a human developmental visual diet yields robust and shape-based AI vision. Nature Machine Intelligence, 8, 735-748. https://doi.org/10.1038/s42256-026-01228-6 https://www.nature.com/articles/s42256-026-01228-6
Preprint version (open access): https://arxiv.org/abs/2507.03168
Supporting References
On the texture bias problem in AI vision — foundational work:
Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., and Brendel, W. (2019). ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. ICLR 2019. https://arxiv.org/abs/1811.12231
On adversarial vulnerability in standard vision models:
Szegedy, C. et al. (2014). Intriguing properties of neural networks. ICLR 2014. https://arxiv.org/abs/1312.6199
On developmental vision in infants — neuroscience foundation:
Courage, M.L. and Adams, R.J. (1990). Visual acuity assessment from birth to three years using the acuity card procedure. Optometry and Vision Science, 67(9), 713-718.
Banks, M.S. and Salapatek, P. (1983). Infant visual perception. In P. Mussen (Ed.), Handbook of Child Psychology, Vol. 2. Wiley.
On the CATDiet related developmental training approach (closely related preprint):
Learning to See Through a Baby’s Eyes: Early Visual Diets Enable Robust Visual Intelligence in Humans and Machines (2025). https://arxiv.org/abs/2511.14440
On scaling laws and the limits of compute-scaling in vision:
Zhai, X. et al. (2022). Scaling vision transformers. CVPR 2022. https://arxiv.org/abs/2106.04560
Hoffmann, J. et al. (2022). Training compute-optimal large language models (Chinchilla). NeurIPS 2022. https://arxiv.org/abs/2203.15556
Coverage and Commentary
ResearchGate preprint version with full PDF: https://www.researchgate.net/publication/393478029_Adopting_a_human_developmental_visual_diet_yields_robust_shape-based_AI_vision
Sam Brown holds a BSc in Electronic Systems Engineering from University of Regina, a Masters from the University of Alberta and a PhD from the University of Toronto in Electrical Engineering. His reporting interests include Physics and Artificial Intelligence.