Curiosity shapes brain-like architectures and functions
This paper demonstrates that combining infant-like curiosity with brain-like biophysical constraints is sufficient to drive the emergence of complex neuronal architectures and cognitive abilities, such as compositional generalization, in artificial neural networks.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
The human mind is a place of endless possibility. From a handful of simple building blocks, we can construct infinite thoughts, plan for distant futures, and make sense of a world that never stops changing. For over a century, scientists have mapped the behaviors and brain circuits that allow this to happen, yet a fundamental mystery remains: how does a system made of simple parts become a mind capable of such profound complexity? One long-standing idea suggests that the brain does not come pre-wired with every answer. Instead, it might self-organize, shaping itself through the interaction of biological limits and the experiences it gathers over time. However, testing this theory in a living human brain is impossible; we cannot strip away a child's curiosity or rearrange their brain's wiring to see what happens. To solve this puzzle, researchers have turned to a different kind of laboratory: the computer simulation. By building artificial brains that follow specific rules, scientists can watch complexity emerge from simplicity, offering a window into how our own minds might have taken shape.
In a new study, researchers set out to see if two specific ingredients were enough to build a mind-like machine. They took a population of artificial neural networks—computer programs designed to mimic the way neurons connect—and gave them two human-like traits. The first was a form of curiosity, modeled after the way infants explore the world. The second was a set of physical constraints, mimicking the biological costs that real brains face, such as the energy required to build long connections between distant parts of the brain. The researchers wanted to know if these two factors, working together, could cause the artificial networks to develop the same complex structures and abilities seen in human brains.
To capture the essence of curiosity, the team first studied 275 infants, aged eight to fifteen months. They watched how the babies looked at visual patterns on a screen, noting exactly when they decided to look away. The researchers found that the infants were not looking randomly. Instead, they behaved like foragers searching for food, but their "food" was new information. The babies stayed engaged with a pattern as long as it was teaching them something new. The moment the pattern stopped offering fresh insights, and their learning slowed down, they would disengage and move on. This behavior fits a principle from ecology known as the marginal value theorem, which predicts that an agent should leave a resource when its local return drops below the average return available in the environment. The researchers translated this rule into a computer algorithm, allowing their artificial networks to decide when to switch tasks based on how much they were learning at that moment.
The second ingredient involved the physical layout of the networks. Real brains are not just a bag of neurons; they are packed into a skull where long connections are expensive to build and maintain. To mimic this, the researchers placed their artificial networks into a virtual space that matched the actual geometry of the human brain. They penalized the networks for creating long-distance connections, forcing them to be efficient, just as evolution has forced the human brain to be. They then trained these networks on a battery of twenty different cognitive tasks, ranging from simple matching games to complex decision-making challenges. The networks were free to choose which task to practice next, guided by their curiosity algorithm, while being constrained by their physical wiring costs.
The results were striking. The networks that combined curiosity with physical constraints did not just learn the tasks; they developed internal structures that looked remarkably like the human brain. Their connections organized themselves into distinct modules, much like the specialized regions of the human brain, and they formed a "rich club" of highly connected hubs, a feature critical for efficient communication in the human mind. These networks also followed a developmental path that mirrored human growth. Just as human brains experience a rapid explosion of connections in early childhood followed by a slow, careful pruning phase that continues into adulthood, these artificial networks showed a similar rise and fall in their connection density. The networks that best matched human brain structure also showed the most human-like pattern of cognitive performance, improving as they "aged" and then gradually declining, just as human abilities do over a lifetime.
Perhaps most importantly, these human-like networks displayed a hallmark of complex human thought: the ability to recombine what they had learned. The researchers tested this by freezing the networks' internal connections and asking them to solve new problems that were mixtures of the old tasks. For example, if a network had learned to avoid a distraction and learned to make a decision, could it instantly solve a new task that required both? The networks driven by curiosity and physical constraints solved these novel combinations rapidly. They did not need to relearn the basics; they simply recombined the internal tools they had already built. In contrast, networks that learned without curiosity, or without the physical constraints, struggled to make these connections.
The study suggests that curiosity is not merely a byproduct of having a complex brain; it is a driver that helps build that complexity. By actively seeking out new information and stopping when it is no longer useful, an agent shapes its own internal architecture. When this curiosity is paired with the physical limits of a biological system, the result is a structure capable of flexible, human-like thinking. The findings offer a new perspective on how the mind emerges, suggesting that the path to complex cognition is not about building bigger or more powerful systems, but about allowing simple constraints and active exploration to unfold over time. This approach provides a testable framework for understanding how the interplay of biology and experience creates the unbounded range of thoughts that define the human experience.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.