Somewhere between the monkey's visual cortex and the silicon of a laptop, a question about efficiency has taken shape. A team led by Ben Cowley at Cold Spring Harbor Laboratory, working with colleagues at Carnegie Mellon and Princeton, has built a machine vision model that borrows its logic from the V4 neurons of macaques, cells that quietly sort colour, texture and shape long before a thought forms. By compressing an enormous set of variables into a model small enough to send by email, they suggest that seeing well may depend less on brute force than on the lessons biology has already worked
Compact AI Model Built From Macaque Visual-Cortex Data Aims to Cut Computing Demands
A model small enough to travel as an email attachment
Mark
What is the headline number here?
Mimi
Sixty million variables compressed to 10,000. Ben Cowley of Cold Spring Harbor says accuracy held up through that reduction.
Mark
And where did the model come from?
Mimi
Data from macaque V4 neurons, which handle colour, texture and shape. Carnegie Mellon and Princeton researchers worked with Cowley's team.
Luke
Before we get excited: that accuracy claim comes from the researchers. The report doesn't say how accuracy was measured or compared.
Mark
So what does 'email-sized' actually tell us?
Mimi
That redundancy was removed using statistical techniques like image compression. It's a vivid way to say the model is small.
Luke
Vivid, yes, but not a benchmark. We don't know its performance against other leading models, or how much energy it saves beyond 'a fraction'.
Mark
What about the self-driving car idea?
Mimi
It's offered as an example of what might become possible, that cars might not need powerful computers to spot pedestrians.
Luke
Note the 'might'. Nothing here says it has been tested on a vehicle. That's a possibility, not a result.
Mark
Why does Chklovskii's comment matter?
Mimi
He's an outside voice at the Flatiron Institute, saying biology-inspired compact AI could lead toward more human-like AI. It's encouraging, though it's an opinion rather than a replication.
The Pulse
- A model that once held 60 million variables has been pared down to 10,000, with the researchers claiming no loss of accuracy.
- The engineering challenge runs against the industry's habit of ever-larger computing demands, pushing against the assumption that better vision requires bigger machines.
- Researchers are stripping out redundancy and borrowing compression techniques from image processing to shrink the system while keeping its perception intact.
- The result is still in the laboratory, with its claims resting largely on the researchers' own description, so independent validation is the next gate before anyone trusts it in a car.
Somewhere between the monkey's visual cortex and the silicon of a laptop, a question about efficiency has taken shape. A team led by Ben Cowley at Cold Spring Harbor Laboratory, working with colleagues at Carnegie Mellon and Princeton, has built a machine vision model that borrows its logic from the V4 neurons of macaques, cells that quietly sort colour, texture and shape long before a thought forms. By compressing an enormous set of variables into a model small enough to send by email, they suggest that seeing well may depend less on brute force than on the lessons biology has already worked out.
A computer model small enough to travel as an email attachment is being offered as proof that machine vision does not have to be enormous. Scientists have built an artificial intelligence system that mimics part of the primate visual system, and they say it needs far less computing power than the models it is meant to rival.
The work comes from the lab of Ben Cowley at Cold Spring Harbor Laboratory. His team began with 60 million variables and compressed them to 10,000, a reduction of several thousand fold, which he says cost no accuracy. The model draws on data from V4 neurons in the macaque visual cortex, cells that handle colour, texture and shape and help primates read their surroundings quickly.
Collaborators at Carnegie Mellon University and Princeton University helped strip out redundancy and apply compression methods borrowed from image processing. The aim is twofold: to understand the brain better and to find ways to improve artificial intelligence. Mitya Chklovskii of the Flatiron Institute suggested that biologically inspired compact AI could move the field toward more human-like systems.
The practical promise is clearest for self-driving cars, which might spot pedestrians without powerful onboard computers. For now, the findings rest largely on the researchers' own account, and independent validation, along with proof that the approach works beyond laboratory data, remains to be seen.
A computer model small enough to travel as an email attachment is now being offered as evidence that machine vision does not have to be enormous. Scientists have built an artificial intelligence system that mimics part of the primate visual system, and they say it needs far less computing power than the AI models it is meant to rival.
The work comes from the lab of Ben Cowley, an assistant professor at Cold Spring Harbor Laboratory. According to Cowley, the model began with 60 million variables. His team compressed them to 10,000, and he says the model lost no accuracy in the process. That is a reduction of several thousand fold.
The model is built on data from V4 neurons, cells in the visual cortex that handle colour, texture and shape. Cowley describes them as responding to things like an arrangement of fruit and the outlines of objects, which is part of how primates read their surroundings quickly and efficiently. The data came from macaque monkeys.
Cowley's team did not work alone. Researchers at Carnegie Mellon University and Princeton University collaborated on the project. To get from the original model to the compact one, they stripped out redundancy and applied statistical techniques similar to those used in image compression. The result is the email-sized system.
The stated purpose is twofold: to help scientists understand more about how the brain works, and to learn how artificial intelligence might be improved. Traditional AI systems need vast amounts of computing power to do comparable visual tasks, while this compact model uses a fraction of the energy.
Mitya Chklovskii, a group leader at the Simons Foundation's Flatiron Institute, said that compact AI inspired by biology could point the way toward more human-like artificial intelligence.
The practical implications are easiest to see in self-driving cars. If a vision system can be this light, such vehicles might not need powerful computers on board to identify pedestrians or obstacles. The model also offers an account of how the human brain recognises faces and objects so fast and so accurately.
The reporting here rests largely on the researchers' own description of their results. What to watch for is independent validation of the compression claim, and whether the approach carries over from laboratory data to real-world uses like driving.
Notable Quotes
Compact, biology-inspired AI could lead the way toward more human-like artificial intelligence.— Mitya Chklovskii, Simons Foundation Flatiron Institute (paraphrased)
Want the full story?
Read the original at The News International ↗
Technology