Francesco Poli at the University of Cambridge and colleagues trained thousands of artificial neural networks with different learning styles to understand how brains become complex. Networks designed with high curiosity, which switched tasks when their learning rate dropped below a threshold, developed more sophisticated cognitive abilities than low-curiosity networks that switched tasks randomly.
The researchers collected eye-tracking data from 275 babies aged 8 to 15 months. Infants were shown sequences of colourful shapes on screens with varying amounts of predictable information. Babies disengaged and looked for something new whenever their rate of learning dropped below average.
All networks, curious or not, became experts at their training tasks. But curious models performed significantly better on tests they had not seen before, demonstrating flexible thinking described as a hallmark of complex human cognition.
When the researchers embedded networks into virtual three-dimensional space modelled on human brain anatomy and penalised long-distance connections to mimic biological costs, curious networks spontaneously developed brain-like features. They formed hierarchical modular structures with densely interconnected areas linked to each other.
The curious networks showed a particular trajectory: connection density rose, then fell, mirroring human brain development. In human brains, synaptic density rises steeply during the first year of life before pruning removes unhelpful connections. Poli speculated that curiosity could relate to selecting which connections to keep.
Researchers mapped networks' developmental trajectories against typical human profiles. The most curious, human-like networks appeared to be at a developmental age of around 20 years, whereas non-curious networks remained at approximately 9 years.
Alireza Modirshanechi at the University of Göttingen called the finding that curiosity enables human-like brain architecture "very surprising" and "a very strong claim." He noted it would be important to verify that non-curious networks don't eventually develop similar features if given more training time. Poli responded that non-curious networks performed as well on basic tasks, suggesting their lack of complex cognition isn't simply slower development.
Poli argued the findings emphasise curiosity's role in education. Forcing children to do one activity for extended periods "might be counterproductive, especially early on", he said. If developing curiosity and cognition is valued, children may benefit from having choices of activities rather than enforced focus.
The results also inform debates around developing flexible artificial intelligence. Poli suggested building curious AIs that learn like humans might prove fruitful, since constraints and limited resources can drive clever solutions. Modirshanechi noted that large language models like ChatGPT demonstrate scaling works, but incorporating curiosity principles might create more human-compatible systems.