Controlling the Monkey Brain: A Breakthrough in Artificial Neural Networks
In a groundbreaking achievement, a team of three scientists from the Massachusetts Institute of Technology (MIT) has successfully used artificial neural networks to create their own control of the monkey cerebral cortex neural activity. The researchers, led by James DiCarlo, the head of the brain and behavioral sciences department at MIT, Pouya Bashivan, a postdoctoral researcher, and Kohitij Kar, have published their findings in the May 2 online edition of Science.
The researchers’ initial goal was to explore how the brain perceives and understands the visual world. To achieve this, they created a calculation model that simulates the brain’s visual cortex. However, simply creating the model was not enough; they wanted to test its accuracy by controlling individual neurons and visual neural networks of neurons. This was a stringent test, as they had to create their own computing model, known as the “controller,” using the model output of another control system, which was the brain of the monkey experiment in neural activity.
The researchers first used the information obtained from the calculation model to create specific images. These images were distinct from natural images, as shown in FIG. They used a depth neural network model synthesis to create these images, which were designed to strongly activate specific neurons in the brain of their choice.
Experimental results showed that these images could indeed strongly activate specific neurons in the brain of their choice. This was a significant success, as it demonstrated the ability of the artificial nervous system to control the activities of the real nervous system.
Experimental Procedure
Over the past few years, DiCarlo and his colleagues have developed a vision system model based on artificial neural networks. Each network consists of nodes or neurons, which are interconnected with different intensity (also known as weights). The researchers have trained the models on a library of more than one million images, with each image tagged with the most prominent object (e.g., an aircraft or a chair, etc.).
The researchers have demonstrated that the activity patterns of these models “neurons” produced by the animal visual cortex are very similar to the reaction of the same image. They designed a neurophysiological experiment with a closed loop: after recording the matching of the model’s neurons and brain sites, they used a synthetic model of a new “controller” image.
The researchers tested the model’s ability to control neurons in the brain by showing them images of animals and comparing them to the reaction of the same image. They found that the model was able to control the activity of target neurons, while maintaining the activity of nearby neurons at a low level. This was a more difficult task, but the researchers were able to achieve it for most neurons testing.
Results
The researchers used these unnatural synthetic image controllers to test whether the ability of the model to predict the reaction of the brain could be applied to new images. The results showed that the model was very accurate, predicting the brain response patterns to 54% of the images evoked.
The researchers also found that the model was able to control the activity of neurons in the visual cortex, with an average of 40% of the natural image response training model than the composite image. This was the first method of control.
Future Applications
The researchers believe that this method has the potential to play a role in the treatment of depression and other mood disorders. They are currently investigating the possibility of extending their model to provide nourishment for the inferior temporal cortex amygdala, which also involves the amygdala emotional processing.
Pouya Bashivan said, “If we have a good neuron model, it can lead to mood or experience different kinds of disorders, we can use the model to control neurons, to help alleviate the imbalance situation.”
Figure 1: Overview of Synthesis Steps
A) Measured by the two control scenario illustrated.
C) Neural control experiment specific steps.
D) Monkey M (black), N (red) and S (blue) neural sites receptive fields.
Where C shows the four steps of neural control experiments:
a large number of labeled training the neural network on nature image, in order to optimize its parameters, and then remains constant;
the ANN “neurons” mapped to each record V4 nerve site;
use synthetic model obtained differentiable “controller” image, or a single-site control groups;
the light emitting image of the specified pattern is applied to the subject’s retina, nerve to measure the level of control sites.
Demonstrate an understanding of the real neural network artificial neural network feasibility.
Experimental Results
The researchers’ study also helps determine the usefulness of these models, which have sparked heated debate whether they can accurately simulate the visual cortex works. James DiCarlo said, “People can understand whether these models have questioned the visual system to give up controversy in the academic sense, we have demonstrated that these models have been strong enough to contribute to an important new application regardless of whether you understand the model. works, In this sense, it has been very useful.”
Aaron Batista, an associate professor of bioengineering at the University of Pittsburgh, said, “They do it really great. Askew like the ideal image suddenly caught the attention of neurons. Neurons suddenly found that it has been looking for stimulation,” This idea is very great, to turn it into a reality far greater. This may be true understanding of the neural network by far the most powerful proof of artificial neural networks.
Conclusion
The researchers’ breakthrough in controlling the monkey brain using artificial neural networks has significant implications for our understanding of the brain and its functions. The study also has potential applications in the treatment of depression and other mood disorders. The researchers believe that their work will contribute to an important new application, regardless of whether we understand the model works.
References