This compact AI model can mimic the human brain
What's the story
Researchers have developed a highly efficient artificial intelligence (AI) model, inspired by the human brain's visual system. The groundbreaking work was published in the journal Nature. The team started with a model that used 60 million variables but managed to compress it into a version that performed nearly as well with just 10,000 variables. "That is incredibly small," said Ben Cowley, an author of the study and an assistant professor at Cold Spring Harbor Laboratory.
Research implications
Potential to unlock secrets of the human brain
The compact AI model not only works efficiently but also mimics the functioning of a living brain. This opens up new avenues for studying neurological disorders such as Alzheimer's disease. Mitya Chklovskii, a group leader at the Simons Foundation's Flatiron Institute, emphasized that if this AI model replicates nature's strategies, it could provide insights into how human brains work.
Visual cognition
Understanding the human visual system
The study also contributes to understanding the human visual system, which processes light into recognizable objects. Cowley has been exploring how AI systems can perform similar tasks and created a model that simulates one part of the visual system with V4 neurons. These cells encode colors, textures, curves, and complex proto-objects. Existing AI systems use deep neural network models for this purpose but require powerful computers and consider a wide range of possibilities.
Neuron activity
How the research was conducted
Cowley's team aimed to compress these large models into a smaller, more compact form. They started with a model trained on data from macaque monkeys and looked for redundant parts while applying statistical techniques such as those used to compress digital photos. The result was a model small enough to be sent as email attachment. This allowed the team to observe what artificial neurons were doing, revealing that some V4 neurons responded specifically to shapes with strong edges and curves.
AI advancement
Implications for self-driving cars
The specialized nature of such V4 neurons could explain how human and other primate brains make sense of what they see sans relying on massive computing power. Cowley said, "If our brains have less complex models and yet can do more than these AI systems, that tells us something about our AI systems." This suggests that self-driving cars could run on less powerful computers while still accurately distinguishing between pedestrians and airborne plastic bags.