Scientists have created a virtual brain network that can predict the behavior of individual neurons in a living brain.
The model is based on a fruit fly鈥檚 visual system, and it offers scientists a way to quickly test ideas on a computer before investing weeks or months in experiments involving actual flies or other lab animals.
鈥淣ow we can start with a guess for how the fly brain might work before anyone has to make an experimental measurement,鈥� says , a group leader at the Janelia Research Campus, a part of the Howard Hughes Medical Institute (HHMI).
The approach, in the journal Nature, also suggests that power-hungry artificial intelligence systems like ChatGPT might consume much less energy if they used some of the computational strategies found in a living brain.
A fruit fly brain is 鈥渟mall and energy efficient,鈥� says , a professor at the University of T眉bingen and an author of the study. 鈥淚t鈥檚 able to do so many computations. It鈥檚 able to fly, it鈥檚 able to walk, it鈥檚 able to detect predators, it鈥檚 able to mate, it鈥檚 able to survive鈥攗sing just 100,000 neurons.鈥�
In contrast, AI systems typically require computers with tens of billions of transistors. Worldwide, these systems as a small country.
鈥淲hen we think about AI right now, the leading charge is to make these systems more power efficient,鈥� says , a computational neuroscientist at Cold Spring Harbor Laboratory who was not involved in the study.
Borrowing strategies from the fruit fly brain might be one way to make that happen, he says.
A model based on biology
The virtual brain network was made possible by more than a decade of intense research on the composition and structure of the fruit fly brain.
Much of this work was done, or funded, by HHMI, which now has maps that show every neuron and every connection in the insect鈥檚 brain.
Turaga, Macke and PhD candidate were part of a team that thought they could use these maps to create a computer model that would behave much like the fruit fly鈥檚 visual system. This system accounts for most of the animal鈥檚 brain.
The team started with the fly鈥檚 connectome, a detailed map of the connections between neurons.
鈥淭hat tells you how information could flow from A to B,鈥� Macke says. 鈥淏ut it doesn鈥檛 tell you which [route] is actually taken by the system.鈥�
Scientists have been able to catch glimpses of the process in the brains of living fruit flies but they have no way to capture the activity of thousands of neurons responding to signals in real time.
鈥淏rains are so complex that I think the only way we will ever be able to understand them is by building accurate models,鈥� Macke says.
In other words, by simulating a brain, or part of a brain, on a computer.
So the team decided to create a model of the brain circuits that allow a fruit fly to detect motion, like the approach of a fast moving hand or fly swatter.
鈥淥ur goal was not to build the world鈥檚 best motion detector, but to find the one that does it the way the fly does.鈥�
The team started with virtual versions of 64 types of neurons, all connected the same way they would be in a fly鈥檚 visual system. Then the network 鈥渨atched鈥� video clips depicting various types of motion.
Finally, an artificial intelligence system was asked to study the activity of neurons as the video clips played.
Ultimately, the approach yielded a model that could predict how every neuron in the artificial network would respond to a particular video. Remarkably, the model also predicted the response of neurons in actual fruit flies that had seen the same videos in earlier studies.
A tool for brain science and AI
Although the paper describing the model has just come out, the model itself has been available for more than a year. Brain scientists have taken note.
鈥淚鈥檓 using this model in my own work,鈥� says Cowley, whose lab studies how the brain responds to external stimuli. He says the model has helped him gauge whether ideas are worth testing in an animal.
Future versions of the model are expected to extend beyond the visual system and to include tasks beyond detecting motion.
鈥淲e now have a plan for how to build whole-brain models of brains that do interesting computations,鈥� Macke says.
Copyright 2024 NPR
