In a first, the 3D reconstruction of a mouse brain links structure to activity.
Let a mouse nose around a house, and it will rapidly find food and form a strategy to return to it without getting caught. Given the same task, an AI would require millions of training examples and consume a boatload of energy and time.
Evolution has crafted the brain to quickly learn and adapt to an ever-changing world. Detailing its algorithms—the ways it processes information as revealed by its structure and wiring—could inspire more advanced AI.
This month, the Machine Intelligence From Cortical Networks (MICrONS) consortium released the most comprehensive map ever assembled of a mammalian brain. The years-long effort painstakingly charted a cubic millimeter of mouse brain—all its cells and connections—and linked this wiring diagram to how the animal sees its world.
Although just the size of a poppyseed, the brain chunk was packed with an “astonishing” 84,000 neurons, half a billion synapses—these are the hubs connecting brain cells—and over 3 miles of neural wiring, wrote Harvard’s Mariela Petkova and Gregor Schuhknecht, who were not involved in the project.
Brain maps are nothing new. Some capture basic anatomy. Others highlight which genes activate as neurons spark with activity. The new dataset, imaged at nanoscale resolution and reconstructed with AI, differs in that it connects the brain’s hardware to how it works.
The project “created the most comprehensive dataset ever assembled that links mammalian brain structure to neuronal functions in an active animal,” wrote Petkova and Schuhknecht.
The new resource could help scientists crack the neural code—the brain’s computational framework. Distilling seemingly random electrical activity into algorithms could illuminate how our brains form memories, perceive the outside world, and make calculated decisions. Similar principles could also inspire future generations of more flexible AI.
“Looking at it [the map] really gives you an awe about the sense of complexity in the brain that is very much akin to looking up at the stars of night,” Forrest Collman at the Allen Institute for Brain Science, who was part of MICrONS, told Nature. The results are “really stunningly beautiful.”
An Enigmatic Machine
The brain is nature’s most prized computational engine.
Although recent AI advances allow algorithms to learn faster or adapt, the squishy three-pound blob in our heads somehow perceives, learns, and memorizes encounters in a flash using far less energy. It then stores important information to guide decision-making in the future.
The brain’s internal wiring is the heart of its computational abilities. Neurons and other brain cells dynamically connect to one another through multiple synapses. New learning alters the wiring by tweaking synaptic strength to form memories and generate thoughts.
Scientists have already found molecules and genes that connect and change these networks across large brain chunks (albeit with low resolution). But deep dive into the brain’s neural connections could yield new insights.
Mapping a whole mouse brain at nanoscale resolution is still technologically challenging. Here, the MICrONS team zeroed in on a poppyseed-sized chunk of the visual cortex. Often dubbed “the seat of higher cognition,” the cortex is the most recently evolved brain structure. It supports some of our most treasured abilities: Logical thinking, decision-making, and perception.
Despite the cortex’s seemingly different functions, previous theoretical studies have suggested there’s a common wiring “blueprint” embedded across the region.
Deciphering this blueprint is like “working out the principles of a combustion engine by looking at many cars—there are different engine models, but the same fundamental mechanics apply,” wrote Petkova and Schuhknecht. For the brain, we’ll need a cellular parts list and an idea of how they work together.
Rebuilding the Brain
The project analyzed a tiny chunk of a mouse’s visual cortex sliced into over 28,000 pieces, each more than a thousand times thinner than a human hair.
The sections were imaged with an electron beam to capture nanoscale structures. AI-based software then stitched individual sections into a 3D recreation of the original brain region, with brain cells, wirings, and synapses each highlighted in differed colors.
The map contains over 200,000 brain cells, half a billion synapses, and more than 5.4 kilometers of neural wiring—roughly one and a half times the length of New York City’s Central Park.
Although it’s just a tiny speck of mouse brain, the map pushes the technological limits for mapping brain connections at scale. Previous landmark maps from a roundworm and fruit fly contained a fraction of the total neurons and synapses included in the new release. The only study comparable in volume mapped the human cortex, but with far fewer identified brain cells and synapses.
Into the Looking Glass
The dataset is unusual because it recorded specific activity from the mouse’s brain before imaging it.
The team showed a mouse multiple videos on a screen, including scenes from The Matrix, as it ran on a treadmill. The mouse’s brain had been genetically altered so that any activated neurons emitted a fluorescent light to mark those cells. Almost 76,000 neurons in the visual cortex sparked to life over multiple sessions. This information was then precisely mapped onto the connectome, highlighting individual activated neurons and charting their networks.
“This is where the study truly breaks new ground,” wrote Petkova and Schuhknecht. Rather than compiling a list of brain components, which only maps anatomy, the dataset also decodes functional connections at unprecedented scale.
Other projects have already made use of the dataset. A few showed how the reconstruction can identify different types of neurons. Mapping structural wiring to activity also revealed a recurring circuit—a generic pattern of brain activity—that occurs throughout the cortex. Using an AI term, the connections formed a sort of “foundation model” of the brain that can generalize, with the ability to predict neural activity in other mice.
The database isn’t perfect. Most of the wiring was reconstructed using AI, a process that leaned heavily on human editing to find errors. Reconstructing larger samples will need further technological improvements to speed up the process.
Then there are fundamental mysteries of the brain that the new brain map can’t solve. Though it offers a way to tally neural components and their wiring, higher level computations—for example, comprehending what you’re seeing—could spark another set of neural activity than that captured in the study. And cortex circuits have vast reach, which means the neural connections in the sample are incomplete.
The consortium is releasing the database, along with a new set of AI-based computational tools to link wiring diagrams to neural activity. Meanwhile, they’re planning to use the technology to map larger portions of the brain.
The release “marks a major leap forwards and offers an invaluable community resource for future discoveries in neuroscience,” such as the basic rules of cognition and memory, wrote Petkova and Schuhknech.
The post Largest Brain Map Ever Reveals Hidden Algorithms of the Mammalian Brain appeared first on SingularityHub.