Curious networks wire themselves like growing brains
Francesco Poli at the University of Cambridge and colleagues tracked the eyes of 275 babies, aged 8 to 15 months, as they watched colourful shapes on a screen 19. Each shape had a learnable pattern for where the next one would appear.
Whenever a baby's rate of learning dropped below their own average, the baby looked away to find something new 19. The team built that rule into thousands of artificial neural networks and trained them on 20 tasks. The high-curiosity networks switched tasks as soon as learning slowed, while the low-curiosity ones switched at random 19. All of them became experts at their training tasks. Only the curious ones did significantly better on tests they had never seen 19. Placed in a virtual 3D space modelled on the human brain, with long-distance wiring penalised, the curious networks formed linked modules of densely connected areas. Their connection density also rose early and then fell as connections were pruned, which is the human pattern 19.