By skyforbes Link: https://arxiv.org/abs/2507.03168
The idea: instead of high-fidelity training from the get-go (the de facto gold standard), we simulate the visual development from newborns to 25 years of age by synthesising decades of developmental vision research into an AI preprocessing pipeline (evelopmental Visual iet – V).
We then test the resulting NNs across a range of conditions, each selected because they are challenging to AI:
- shape-texture bias
- recognising abstract shapes embedded in complex backgrounds
- robustness to image perturbations
- adversarial robustness.
We report a new SOTA on shape-bias (reaching human level), outperform AI foundation models in terms of abstract shape recognition, show better alignment with human behaviour upon image degradations, and improved robustness to adversarial noise – all with this one preprocessing trick.
This is observed across all conditions tested, and generalises across training datasets and multiple model architectures.
We are excited about this, because V may offers a resource-efficient path toward safer, perhaps more human-aligned AI vision. This work suggests that biology, neuroscience, and psychology have much to offer in guiding the next generation of artificial intelligence.