A Drosophila Whole-Connectome Network Can Learn Human-Designed Cognitive Tasks
This study demonstrates that the fixed anatomical wiring of a male *Drosophila* connectome serves as an effective computational substrate for learning human-designed cognitive tasks, such as bounded addition and grounded relational language, significantly outperforming randomized network rewires.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The human brain is often described as the most complex object in the known universe, a tangled web of billions of cells that allows us to think, move, and speak. For decades, scientists have tried to map this wiring, creating detailed diagrams called connectomes that show exactly which neurons are connected to which. These maps are usually studied to understand how the brain controls the body, such as how a fly sees a flower or how a human lifts a cup. But a new line of inquiry asks a bolder question: if you take the physical wiring diagram of a brain and use it as the skeleton for a computer program, can that program learn things the animal never evolved to do? It is a test of whether the specific shape of biological connections holds a secret power that random connections do not.
Researchers at the Korea Advanced Institute of Science and Technology decided to test this idea using the complete nervous system of a single male fruit fly. They took the publicly available map of this fly's brain, which contains over 166,000 neurons and nearly 26 million connections, and used it as the fixed structure for an artificial network. They did not try to simulate the fly's actual thoughts or movements. Instead, they trained this biological skeleton to solve two tasks that have nothing to do with fruit flies: adding numbers and understanding a simple language about sensory events. In the first task, the network had to learn to add two numbers together, where the sum never exceeded one hundred. In the second, it had to listen to a description of two physical sensations, like a strong pressure on a leg followed by a weak movement, and then correctly match that description to the event, even when the order of the events was reversed.
To see if the fly's specific wiring mattered, the scientists created a control group. They took the same number of neurons and connections but shuffled the links around randomly, keeping only the number of connections each neuron had the same. This created a "rewired" brain that looked statistically similar but lacked the specific, evolved pathways of the real fly. When they trained the real fly network on the addition task, it succeeded with high accuracy, getting the right answer nearly 93 percent of the time on new problems it had never seen before. The random, rewired networks struggled significantly, managing only about 68 percent accuracy. The difference was substantial, suggesting that the specific arrangement of the fly's brain provided a hidden advantage for learning math, even though the fly itself never does math.
The results were even more striking with the language task. Here, the network had to connect sensory inputs to words, learning that a specific pattern of signals meant "strong pressure" and that this happened "before" a "weak movement." The researchers tested the network in four different ways, changing where the sensory information entered the brain and where the words came out, to ensure the results weren't just a fluke of one specific setup. Across all these tests, the network built on the real fly brain consistently outperformed the rewired versions. On the most difficult test, which required the network to correctly pair a scene with its description and also the reversed scene with its reversed description, the real brain achieved a score of roughly 62 percent, while the random versions averaged only 44 percent. In a larger comparison involving twenty different random rewires, the real fly brain ranked first, beating every single one of them.
Crucially, the study showed that this advantage was not just about having the right number of connections. The researchers found that the real brain's success depended heavily on the alignment between the sensory inputs and the specific parts of the brain they were connected to. When they scrambled the sensory signals so that the input no longer matched the brain's natural wiring, the performance of the real brain collapsed, dropping from about 61 percent to just 19 percent. This proved that the brain's structure was not just a generic container for learning; it was a specialized machine where the physical location of the inputs mattered deeply. The random networks, lacking this specific alignment, could not recover from the confusion.
The findings suggest that the way a brain is wired contains a reusable blueprint for learning, an inductive bias that helps it solve problems efficiently. This does not mean the fly brain is a universal computer that can do anything. The network still needed to be trained from scratch, and it could not learn without the correct sensory inputs. However, the study demonstrates that the biological arrangement of edges in a connectome is more than just a record of an animal's past; it is a functional architecture that can be repurposed for human-designed cognitive tasks. The specific pathways that evolved for a fly to navigate its world also happen to provide a head start for a computer trying to learn arithmetic and language, a discovery that challenges the idea that biological wiring is only useful for the behaviors it originally evolved to perform.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.