Scientists harness monkey neurons to create pocket-sized AI brain: NPR

The human brain uses less power than a light bulb, but artificial intelligence systems use a lot more power to perform the same tasks.

Now, scientists have created a highly efficient AI model that suggests how the living brain can do more with less, a team reports in the journal nature.

The model, which mimics part of the brain’s visual system, started with 60 million variables. However, the team was able to compress it into a nearly equally performing version using just 10,000 variables.

“This is incredibly small,” says study author Ben Cawley, an assistant professor at Cold Spring Harbor Laboratory. “This is something you can tweet or email.”

The compact model also appears to function more like a living brain, which could help scientists study what goes wrong in diseases such as Alzheimer’s disease, Cowley said.

More broadly, if AI models truly replicate strategies seen in nature, they could help scientists understand the inner workings of the human brain, said Mitya Chiklovsky, group leader at the Simmons Foundation Flatiron Institute, who was not involved in the study.

Compact, biologically-inspired brain models could also lead to “more powerful, more human-like artificial intelligence,” said Chiklovsky, who is also a faculty member at New York University Medical Center.

monkey data

The research is part of an effort to understand the human visual system, which takes fragments of light and transforms them into things we can recognize, like Grandma or the Grand Canyon.

Cowley said scientists studying the visual system are trying to answer questions like “How do we recognize cats?” or “How do you tell a dog apart?”

There’s no good way to watch the human brain do this. So Cowley has turned his attention to artificial intelligence systems that can perform similar tasks.

But there’s a problem. “Just like our own brains, we have a very poor understanding of how these AI systems work,” Cowley says.

Cawley worked with researchers at Carnegie Mellon University and Princeton University to create an AI model that the team could understand. It simulates only a part of the visual system, which is characterized by cells called V4 neurons.

“They encode colors, textures, curves, and very complex proto-objects,” Cowley says.

Existing AI systems can do the same thing using deep neural network models, but this requires powerful computers and learns over a huge range of possibilities. But Cowley’s team wanted something more efficient.

“We’re looking to compress these big, clunky models into much smaller, more compact forms,” he says.

They started with a model trained on macaque data. Next, I looked for redundant or unnecessary parts of the model. Statistical techniques, such as those used to compress digital photographs, were also applied.

The result is a model small enough to be attached to an email.

Compact model with fewer secrets

The model was so small and simple that the researchers were able to get a glimpse of what the artificial neurons were doing.

For example, some V4 neurons responded to shapes with strong edges and many curves (the kind of shapes you often see in the produce aisle at the grocery store).

“When you go to the supermarket and see fruit on display, your V4 neurons appreciate it,” Cowley says. “They love arranged fruit. They love all the curves in the apple.” [and] orange. ”

Other V4 neurons seemed to respond only to small points in the image.

“This was really interesting to us because primates are very visual animals,” Cowley said.

The special properties of these V4 neurons may help explain how the brains of humans and other primates are able to understand what they see without relying on massive computing power.

This discovery could also have implications for artificial intelligence.

“If our brain models are less complex and can still do more than these AI systems, that tells us something about our AI systems,” Cowley says. That is, they are probably smaller and simpler, yet may be able to better interpret what they see.

For example, a self-driving car might be able to correctly distinguish between a pedestrian and a plastic bag in the air, while still being able to navigate on less powerful computers, he said.

But for AI systems to perform as well as the human brain, Chklovskii says they need to do more than just shrink.

For example, he says, people can easily recognize a friend’s face in any situation and from different angles, even if that friend has a tan or a new hairstyle.

AI systems, even those with supercomputers, struggle with these types of tasks.

That may be because current AI models are based on 20th century understanding of the human brain, Chklovskii says.

“Since then, we’ve learned a lot more about the brain,” he says. “Therefore, the foundations of artificial networks may need to be updated.”