Scientists have made a pocket-sized AI brain, inspired by monkey neurons, that's incredibly efficient. This breakthrough could help us understand how living brains work and potentially lead to more powerful, human-like AI. The key? Compressing an AI vision model to just 10,000 variables, making it small enough to send in an email.
The human brain is incredibly energy-efficient, using less power than a light bulb to perform complex tasks. Now, researchers have created an AI model that mimics this efficiency, using data from macaque monkeys. This model, which simulates a part of the visual system, started with 60 million variables but was compressed to just 10,000, performing nearly as well.
This compact model, according to Ben Cowley, an assistant professor at Cold Spring Harbor Laboratory, could be shared in a tweet or email. It also works more like a living brain, which could help scientists study diseases like Alzheimer's. Mitya Chklovskii, a group leader at the Simons Foundation's Flatiron Institute, suggests that this biology-inspired approach could lead to more powerful and human-like AI.
The study, part of an effort to understand the human visual system, focuses on how we recognize objects like cats and dogs. While human brains are great at this, AI systems struggle with similar tasks, lacking a deep understanding of their own workings. Cowley's team created a more efficient AI model by simulating V4 neurons, which encode colors, textures, and shapes.
By compressing the model and applying statistical techniques, they achieved a smaller, more efficient version. This allowed them to observe how artificial neurons respond to specific shapes and patterns, like arranged fruit in a supermarket. Some V4 neurons respond to small dots, possibly explaining why primates are drawn to eyes.
The findings suggest that our brains use less complex models to achieve more than AI systems. This could lead to smaller, more efficient AI, like self-driving cars that can distinguish pedestrians from plastic bags. However, AI still needs to match the human brain's versatility, recognizing faces in various settings and angles, a challenge current models struggle with.
