Neuron Yang Guangyu's team develops artificial neural network to simulate the brain to recognize odors

Editor-in-Chief | Xi October 6, 2021, the team of Guangyu Yang, associate researcher, Department of Brain and Cognitive Sciences, and Assistant Professor of Department of Electrical Engineering and Computer Science at the McGovern Institute, Massachusetts Institute of Technology (the first author is Dr.
Yiliu Wang) on ​​Neuron Published an article Evolving the olfactory system with machine learning, using deep learning to develop an artificial neural network.
After learning to recognize odors, the network can evolve a structure that is highly similar to the olfactory circuit of the brain
.

The brain's olfactory circuit has been most accurately mapped in Drosophila: starting from the antennae, each sensory neuron is equipped with a receptor that detects a specific odor
.

The receptor generates an electrical signal after binding to the odor molecule
.

When an odor is detected, these neurons that make up the first layer of the olfactory network send signals to the second layer of neurons located in the antennal lobe
.

Sensory neurons with the same receptor will connect to the same second-layer neuron
.

Because there are fewer neurons than the first layer, the second layer is considered a compression layer
.

Next, the neurons in the second layer send signals to a larger group of neurons in the third layer—the strange thing is that the connections between the neurons in the two layers seem to be random
.

Through close cooperation with Columbia University neuroscientists Richard Axel and Larry Abbott, Guangyu Yang and Yiliu Wang constructed an artificial neural network, which is a computing tool inspired by the brain, including an input layer, a compression layer, and an expansion layer.
Like the olfactory system of fruit flies
.

The researchers only gave each layer the same number of neurons as the Drosophila system, not a specific structure
.

Artificial neurons in the network master specific tasks by changing the strength of their interconnections
.

Through training, they can recognize patterns in complex data sets, which makes them valuable for speech and image recognition and other forms of artificial intelligence
.

Yang Guangyu said: "This is what the brain's olfactory system is particularly good at
.

If you combine two different apple scents, the brain will still smell the smell of apples
.

But if you superimpose the two cat photos pixel by pixel, the brain will not see a cat
.

this capability is just one characteristic odor processing circuits of the brain, but it captures the essence of the system "
.

Wang Yiliu said: “Networks of different structures can produce similar neural activity, so neuroscientists still need to know whether artificial networks reflect the actual structure of biological circuits
.

With the comprehensive anatomical data of the olfactory circuit in Drosophila, we can ask this question: artificial Can neural networks really be used to study the brain?" Researchers use the network to assign data representing different odors to their respective categories-not only to correctly distinguish a single odor, but also to correctly identify a mixture of odors
.

The artificial network only takes a few minutes to learn this task
.

The structure that emerges is surprisingly similar to the structure in the brain of the fruit fly
.

Each neuron in the compression layer receives signals from the same type of sensory neuron and connects to multiple neurons in the expansion layer seemingly randomly
.

What’s more interesting is that each neuron in the expansion layer receives an average of connections from 6 neurons in the compression layer—just like what happens in the fly’s brain
.

"This number can be 1, it can be 50, or it can be any number in between," Yang Guangyu said: "We found 6 in living organisms, and our network roughly found 6 too
.
"
Biological evolution has discovered this organizational principle through random mutation and natural selection, while artificial networks have found it through standard machine learning algorithms
.

This surprising convergence provides strong support for the hypothesis that the brain has the optimal structure for analyzing olfactory information
.

Now, researchers can use the model to further explore this structure: to explore how the network evolves under different conditions, and to manipulate loops in ways that experiments can't be done
.

In short, this work can not only help researchers better understand the brain's olfactory circuit, but also help embody the connection between artificial neural networks and neuroscience
.

When people see that artificial neural networks can accurately match the brain architecture, they will be more convinced that neural networks will have a lot to do in brain modeling
.

Original link: https:// Platemaker: Instructions for reprinting on the 11th [Non-original article] The copyright of this article belongs to the author of the article, and personal forwarding and sharing are welcome.
Reprinting is prohibited without permission, the author has all legal rights, and offenders must be investigated
.