This network of artificial neurons affects the visual cortex in a real neural network

"That they managed to do that is really amazing."

The researchers created a network of artificial neurons that helps them better understand the operation and manipulation of real neural networks.

MIT researchers James DiCarlo, Pouya Bashivan and Kohitij Kar created a computer model that allowed them to create images that could then be used to strongly influence specific neurons in the brain during animal testing.

For several years, researchers have designed models of the visual cortex and the visual system. To create these artificial neural networks, they started by creating an arbitrary architecture composed of nodes representing individual neurons. These nodes can connect with more or less force.

These patterns were then used to power a library of over a million images, each containing a label indicating the most important object in the image, such as a car or a type of food. The network of artificial neurons entails learning what each image is by manipulating the strength of the connection between each node.

The researchers found that the nodes of the artificial neural network reacted in a very similar way to the reaction of real neurons in the animal's visual cortex when the same image was presented to them.

Pouya Bashivan, one of the leading authors of the article, commented in a blog post from MIT: "These models have predicted what would be the neuronal responses to other stimuli that are noticeable in the brain. they had never seen it before. "

The research could help neuroscientists understand the interaction of neurons and thus lead to new treatments for neurological disorders such as Alzheimer's disease, epilepsy or depression.

Manipulation of real neural networks

The team then deepened the research and verified whether the models they had created could be used to create images that could manipulate a real neuron into a "desired state." Essentially, they wanted to test whether the model could control the activity of neurons in the visual cortex of an animal.

To test this, the team created an individual map of a specific part of an animal's brain. The selected area was V4, which contains the visual cortex and houses millions of neurons. The researchers mapped five to 40 neurons at a time. This was done by showing both the animals and the computer model, images from the library, and then recording and comparing subsequent neuronal response patterns.

James DiCarlo, head of MIT's Department of Brain Science and Cognition, said, "Once every neuron has a mission, the model allows you to make predictions about that neuron."

In one test, they used the artificial neural network to create synthetic images that did not look like natural objects, but still failed to respond to target neurons. When these synthetic images were shown to the experimental subject, the targeted neurons responded 40% of the time.

"The fact that they managed to do that is really amazing. It is as if, for this neuron at least, his ideal image suddenly became clear. The neuron suddenly introduced the stimulus he had always sought, "commented independent critic Aaron Batista, an badociate professor of bioengineering at the University of Pittsburgh.

"It's a great idea, and making it a reality is quite an achievement. This is perhaps the strongest validation to date of using artificial neural networks to understand real neural networks, "said Batista.

Researchers are working to improve the model by integrating new data from the use of computer-generated images.

"If we had a good model of neurons involved in emotions or causing various types of disorders, we could use this model to drive them to help improve these disorders," Bashivan concluded.

See also: Search Teams Use Summit Computer to Win Gordon Bell Award