Why brains learn similarly early on, but diverge later

Researchers tested five learning algorithms (including backpropagation and predictive coding) against fMRI from humans and neural recordings from macaque monkeys, finding that early visual areas show conserved alignment across species but higher areas do not. A pretrained ImageNet model dramatically outperformed custom networks at macaque IT, suggesting that brain alignment downstream depends more on training quality than on which learning rule the model uses. The findings split visual cortex into two functional zones: a conserved early stage and a flexible higher stage shaped by task demands.