Curiosity Shapes Brain-Like AI Development — SkimNews
Get the Health newsletter
Daily health & science — research, biotech, public health, the studies worth knowing. Free.
- Francesco Poli and colleagues trained artificial neural networks with varying levels of curiosity, finding that high-curiosity models developed more flexible cognition and human-like brain architecture
- Infant eye-tracking data from 275 babies between 8 and 15 months revealed that infants disengage when learning rates drop, a behavior used to model curiosity in artificial networks
- Curious neural networks spontaneously formed modular, hierarchical structures and followed a developmental trajectory mirroring human brains, including early overproduction and later pruning of connections
- Virtual 3D brain models showed that only when curiosity was combined with wiring efficiency constraints did networks develop complex, anatomically plausible structures resembling real brains
- High-curiosity networks performed significantly better on novel, unseen tasks after training, demonstrating a hallmark of human-like cognitive flexibility despite equal task mastery during training
- Alireza Modirshanechi called the emergence of brain-like structure from curiosity 'very surprising' and emphasized the need to test whether non-curious networks could achieve similar complexity over longer training periods
Why it matters: This research shifts the understanding of curiosity from a byproduct to a driver of brain complexity, with implications for education and AI development. If curiosity fundamentally shapes cognitive architecture, then learning environments and artificial systems that suppress or neglect intrinsic exploration may fail to cultivate advanced reasoning, limiting both human potential and machine adaptability.
Ask SkimNews
