Scientists have been able to thoroughly pump computer vision thanks to the brain of monkeys
The computer model of vision became more resistant to image distortions when it began to imitate the work of biological neurons.
Computer vision is the ability of artificial intelligence to analyze visual information, such as images from cameras or sensors in self-driving cars. However, computer vision does not always accurately recognize objects. Scientists at the Massachusetts Institute of Technology (MIT) and companies IBM found a way to improve computer vision by making it mimic how a real, living brain works.
Artificial intelligence technology is based on neural networks – mathematical models consisting of many interconnected “neurons” that process information. They are somewhat similar to the biological neural networks in the brain of humans and other primates that are responsible for visual perception, but much less functional than real ones.
Scientists have found that when neural networks effectively solve computer vision problems, they form strong neural circuits that work in a very similar way to the same neural circuits in the living brain.
However, computer vision systems are still completely incomparable with human vision. James DiCarlo, a professor at MIT, has suggested that one way to improve computer vision could be to incorporate specific features of the human brain into these models.
To test this idea, DiCarlo and his colleagues used data from vision-processing neurons in the lower temporal cortex of monkeys that were obtained while showing various images to primates. The scientists then had the neural network mimic the behavior of those neurons while the network learned to recognize objects in the images.
After training the artificial model with biological data, DiCarlo’s team compared its activity with a similar neural network model trained without the aforementioned data. The scientists found that the new model matched data from neurons in the lower temporal cortex much better and was much more similar to human vision in terms of how it works. It was also more resistant to “adversary attacks” – small distortions in images that usually mislead neural networks.
The new work is further evidence that the exchange of ideas between neuroscience and computer science can drive progress in both fields. “Everyone benefits from the exciting cycle of interaction between natural and artificial intelligence,” said DiCarlo.