正文 · AI 翻译
IT Home reported on November 2: Why did the human brain evolve to be so complex? The answer may come down to curiosity. Researchers trained thousands of artificial neural networks with different learning patterns and found that the most curious networks not only had more complex and flexible cognitive abilities, but also spontaneously developed structures resembling the human brain.
These networks formed interconnected modules and followed developmental patterns similar to those of humans: for example, they first developed many dense neural connections, then gradually pruned them. Francesco Poli of the University of Cambridge said that curiosity is usually seen as a result of a complex brain. “But could curiosity, in turn, have shaped the brain’s form?”
To simulate how infants learn, Poli and his colleagues collected eye-tracking data from 275 infants aged 8 to 15 months. The researchers displayed a series of colorful shapes on a screen. The sequences conveyed varying amounts of information, offering clues about where the next shape would appear. Poli said, “The positions were not random; each set of shapes had a pattern that could be learned.”
The experiment found that infants were highly sensitive to how much information different stimuli carried. Whenever their rate of learning fell below average, they lost interest, looked away, and sought something new.
Based on this principle, the research team built artificial neural networks with different levels of intrinsic curiosity and had them complete 20 tasks. Highly curious models were set to switch tasks immediately once their learning rate fell below a threshold; less curious models switched tasks at random.
All the networks mastered the training tasks, whether or not they had curiosity. After training, however, the curious models performed markedly better on new tests they had never encountered. Poli said this ability to flexibly combine different information is a hallmark of complex human cognition.
The curious models also developed structural features resembling those of complex biological brains. The researchers embedded the artificial networks in a three-dimensional virtual space modeled on human brain anatomy and penalized long-distance connections to simulate the high metabolic cost of long-range neural wiring in real brains, pushing the networks to form efficient connections.
That constraint alone was not enough to produce complex structures. But when the curiosity mechanism was added, a series of brain-like features emerged. For example, curious networks developed a hierarchical modular structure: functional blocks with dense internal connections were also connected to one another.
Poli also observed a developmental pattern that surprised him: the density of network connections rose and then fell. He said, “I thought it was very strange at the time.”
But this closely matches human brain development: during the first year after birth, the number of synapses that transmit signals between neurons surges, followed by a pruning phase in which unused connections are removed. He speculated that curiosity drives individuals to actively seek learning tasks offering the greatest information gain, and that this mechanism “may be related to the process of selecting which synapses to keep or discard.”
The researchers compared changes in network connection density with a standard curve of human brain development to estimate the “physiological age” corresponding to each network. The results showed that the most curious networks, whose structures most closely resembled the human brain, were at a more mature developmental stage, corresponding to an age of about 20 years; networks without curiosity, whose wiring patterns were off, remained at an earlier stage, corresponding to an age of about 9 years.
Alireza Modirshanechi of the University of Göttingen in Germany said it was unsurprising that curiosity improves learning efficiency: it keeps us from wasting effort on things that are too hard or too easy. “You prioritize learning the things that are most suitable for learning.”
But the study’s suggestion that curiosity can give rise to brain-like architecture was “very surprising,” he said. “If true, this is a weighty claim.” He believes it still needs to be tested whether networks without curiosity might simply develop more slowly and eventually form similar structures if trained for long enough.
Poli’s view is that networks without curiosity performed just as well as curious networks on the basic training tasks, so their failure to develop complex cognition and brain-like structures was not simply a matter of slower learning. “I think that even with longer training, they would remain largely the same. What curiosity really changes is how the networks learn.”
Even so, Poli said the findings also highlight the value of curiosity in education. An environment that forces children to spend long periods doing the same thing “may be counterproductive, especially in early childhood.” “It depends on what we want to cultivate. If we value the ability to focus on a single task for a long time, then we need to train that ability specifically.” But if we value curiosity and cognitive ability, we should first offer children a variety of activities to choose from, “then let them explore independently within that range.”
IT Home noted that the study also contributes to the discussion about the direction of development for flexible, human-like artificial intelligence. Poli said the vast majority of AI research simply “makes systems bigger,” consuming ever more data and computing power. He suggested that building curiosity-driven AI with learning patterns closer to those of humans may be a viable path to complex cognition. Evolution itself did not have unlimited resources, and resource constraints are precisely what gave rise to ingenious solutions. “Sometimes, these constraints are exactly why we become good at what we do.”
Modirshanechi, however, does not think AI developers must change direction. He said large language models such as ChatGPT have already shown that scaling works. But incorporating principles of curiosity could help create systems better suited to human needs: for example, a model might proactively ask questions to clarify what a user really wants before producing a lengthy answer.
References:
https://www.biorxiv.org/content/10.64898/2026.07.02.735826v2
Advertising statement: External links in this article (including but not limited to hyperlinks, QR codes, passcodes, and other forms) are provided to convey more information and save time in selecting information. The results are for reference only. All IT Home articles containing external links include this statement.