Generic fitting models learn edge representations from prenatal retinal waves.

Orientation selectivity-the representation of oriented edges-is a hallmark of biological vision, shared across mammals, birds, and reptiles. However, the origins of orientation selectivity are unknown. Is orientation selectivity predetermined, with genes instructing the development of edge representations? Or is orientation selectivity the product of blind evolution-like (variation + selection) fitting during prenatal development? Here, we provide evidence supporting the fitting account. Using generic image-computable fitting models (transformers), we show that orientation selectivity develops when fitting systems adapt to prenatal experiences. Our models started from scratch, with no innate orientation selectivity and no hardcoded priors about lines, objects, or space. The models were then trained with a biologically plausible fitting objective (unsupervised temporal learning) and biologically plausible prenatal data (retinal waves). Despite starting from scratch, the models spontaneously developed robust orientation selectivity. This result generalized across architecture sizes, training conditions, and retinal waves from different species. Edge representations develop when domain-general fitting mechanisms adapt to prenatal experiences, supporting fitting theories of learning and development.