This nematode is not simple! The brain has been restored with high precision and can move forward dynamically

The highest precision "nematode brain", it's here.

This "brain" simulates the entire biological nervous system of a C. elegans worm.

(Note: Caenorhabditis elegans is the "simplest living intelligence", with 302 neurons)

This time, domestic scholars not only combined the entire neuron network of Caenorhabditis elegans It was restored, down to their sub-cellular level connections.

It is understood that its level of sophistication has reached the highest level currently known:

Previously, a study studied the computational complexity of a single biological neuron. The article pointed out that, A deep neural network requires 5 to 8 layers of interconnected neurons to represent the complexity of a single biological neuron.

And through such a carefully constructed "brain", this "intelligent nematode" can complete dynamic crawling.

This is the latest research result from Beijing Zhiyuan Artificial Intelligence Research Institute, and the "sharp weapon" behind it is the Tianyan project.

And the birth of this "intelligent nematode" MetaWorm 1.0 is not only a breakthrough in the accuracy of life simulation, according to the research team:

This is a step forward A critical step in artificial intelligence life.

How is the most precise brain made?

The Caenorhabditis elegans selected for this brain can be said to be "the most precise worm with a nervous system" "One of the simplest organisms" -

It has a complete nervous system, can sense, escape, forage and mate, and its overall structure is very simple. Adult insects only have about 1,000 somatic cells.

This small transparent creature, about 1mm long, has become a "frequent visitor" in the scientific research community. In the past 20 years, three Nobel Prizes have been related to it.

For neuroscientists, the nervous system of Caenorhabditis elegans has been completely cracked, and the real-time map was even on the cover of Nature that year, making it very suitable for research. and simulate “brain circuits.”

△Hermaphrodite, with a total of 302 nerve cells

More importantly, the neurotransmitters such as acetylcholine and dopamine present in nematodes are not found in mammals Also exists.

Studying its nervous system also plays an important role in studying the regulatory mechanism of the human nervous system.

But studying structures is one thing, modeling them with computers is another.

You must know that simulating a biological neuron is not simply a linear transformation like convolution. It simulates the exchange of materials (such as between cells) and the generation of action potentials between neurons. Behaviors such as conduction and conduction are very complex.

For example, just the transmission of transmitters between synapses involves multiple parameters such as quantity, speed, concentration, backflow, direction, etc. It will be more complicated to calculate and simulate using mathematical models.

Even if a complete nervous system is simulated, how to use computers to simulate a "cyberspace" close to the real environment and train an "intelligent nematode" model in it is another major research difficulty.

Previously, although many teams have been conducting research on nematode simulation, there is a certain gap between the accuracy and the simulation environment and reality. For example, our common bionic robot fish is far from the fish. The accuracy is the same.

This time, the Tianyan team successfully modeled the highest-precision intelligent "cybernematode", allowing it to dynamically wriggle forward in a 3D fluid simulation environment and have the ability to simply seek advantages and avoid disadvantages. ability.

So, what does this "intelligent nematode" look like?

First, the team used a large number of formulas and models to model the "electronic neurons" of the nematode.

There are three main types of models used: multiple ion channel models, Hodgkin-Huxley models and multi-compartment models (Multi-compartment Model).

Among them, the multiple ion channel models, as the name suggests, are used to simulate various ion channels on the cell membrane. The Trimble 1.0 model uses 14 kinds of ion channels;

Hodgkin-Huxley model (HH model) , can simulate each part of the neuron into a different circuit component;

△HH model example, the picture comes from Wikipedia - True·biology is a sophisticated electronic instrument

Multi-chamber model, the neuron is regarded as a system, divided into several according to the dynamic characteristics Chambers, each chamber contains a different number of ion channels.

△The picture comes from the paper "Analysis of multi-compartment model of medium spiny neurons" written by Jiang Xiaofang, Liu Shenquan, and Zhang Xuchen

Combination of these three models Together, it is possible to simulate the structure of neurons, the formation and conduction of action potentials and gradient potentials on neuron cell membranes, and the rate of material conduction between various body parts.

After the construction was completed, this "intelligent nematode" carefully modeled the 302 neurons of Caenorhabditis elegans (hermaphrodites) and thousands of connections between these neurons, using 14 types of Ion channels, detailed to the subcellular level.

The 302 neurons of C. elegans are divided into sensory neurons, interneurons and motor neurons. Among them, the team conducted high-precision modeling of 106 sensory and motor neurons, which are highly simulated. combined with their electrophysiological dynamics.

According to statistics, the maximum number of compartments for a single neuron is 2313, and the minimum is 10. 302 neurons averaged 52 compartments each. The synaptic connections between neurons are as fine as the level of neurites (dendrites, axons):

Then, the team constructed a 3D fluid dynamic simulation environment to allow C. elegans moves in a realistic scene.

Note that the step of simulating the environment is particularly important. It is a key step in studying how nematodes adapt to the microenvironmental movement.

After nematode modeling has been refined to the subcellular level (micron level), the scale of physical laws has shrunk, and the effects of friction and viscosity are several orders of magnitude greater than gravity.

In this case, the nematode can still eat and drink water freely to provide energy, which is inseparable from its ingenious way of interacting with the environment.

Therefore, the Tianyan team combined computational neurology, motion mechanics, graphics and other interdisciplinary disciplines to construct a realistic nematode muscle and body software model for the intelligent nematode "Tianbao", and established a more suitable artificial body Fluid simulation environment for training.

Specifically, this environmental framework consists of multiple modules including three-dimensional modeling, finite element solving, simplified fluid model, reinforcement learning, visualization, etc., which can simulate the interaction between nematodes and the environment to the greatest extent.

Compared with the current internationally leading OpenWorm nematode simulation project, the fluid simulation environment of the Tianyan team is larger in scale and is more suitable as a multi-body/swarm intelligent behavior simulation environment for living organisms, completing various intelligent energy bodies. Learn and train complex tasks, etc.

Finally, the team placed the nematode model into the simulation environment and completed preliminary training.

These are all components of the future Tianyan platform. Specifically, this is a multi-GPU cluster platform still under construction that can be used for high-precision, large-scale biological neuron simulation.

In a simulation environment with a scene scale of more than 1,300 nematode lengths, the team has now preliminarily trained "intelligent nematodes" that can act autonomously according to the distribution of environmental chemical signals, and this scene can also support larger spaces and multiple nematode population simulations.

According to the team, the "intelligent nematode" model can efficiently and accurately calculate the rules of interaction with the fluid environment. Under the same computing resources, the single simulation time of a single nematode is less than 0.1 seconds.

In the next stage, the Tianyan team plans to allow this "cyberworm" to achieve more complex intelligent tasks such as obstacle avoidance and foraging.

In fact, brain-like intelligence research has always been a global issue.

Internationally, including the Blue Brain project supported by the European Union Brain Project and the American Brain Project, etc., are conducting brain-like research; technology giants such as Google have been releasing brain maps and brain tools in the past five years; university research institutions For example, MIT used 19 nematode simulated neurons to realize automatic driving control...

However, from the perspective of brain-inspired research alone, the research directions of each team are very different, and a considerable number of teams even borrowed Brain-inspired computing is realized by designing the chip first and then designing the algorithm.

However, such research will constrain the design and implementation of algorithms by hardware such as chips, and ultimately fall far short of the goal of realizing brain-like intelligence.

In contrast, the Tianyan team chose to study and realize brain-like intelligence from the perspective of realizing AI.

But even so, is it really meaningful to go to all the trouble to model a nematode brain?

What is the use of the nematode "brain"?

If you use one sentence To summarize this problem, it is:

This is a key step towards artificial intelligent life.

Since the birth of artificial intelligence, "making machines like humans" has become the direction that researchers have been working hard to develop.

However, with the passage of time, even at the current development stage dominated by deep learning, artificial intelligence still has not reached the level of intelligence in the true sense.

Even the Go game like AlphaGo that shocked the world in 2016 only refreshed people's understanding of artificial intelligence.

But as CMU professor Hans Moravec said:

It is relatively easy to make a computer play chess like an adult; The ability to perceive and act at the level of a one-year-old child is quite difficult or even impossible.

So, where does the problem lie?

In 2016, Huang Tiejun, president of Zhiyuan Research Institute, gave the answer.

He believes that deep learning essentially relies on artificial neural networks, and biological intelligence relies on biological neural networks.

Among them, artificial neural networks are closer to "realizing functions", while biological neural networks simulate "structures that realize functions". The two are not at the same level in terms of "volume", and the latter It is obviously much larger and more important - because structure determines function, and biological neural networks are the carrier of intelligence.

Therefore, the "solution" proposed by Huang Tiejun based on this situation is:

From the perspective of brain mechanism simulation.

To put it simply, it is to explore the "operating mode" inside the biological brain. This is one of the ways to lead to general artificial intelligence.

Coincidentally, earlier in 2009, Professor Henry Markram of the Ecole Polytechnique Fédérale de Lausanne in Switzerland also put forward a similar point of view.

At that time, he announced a plan to use supercomputers to build a brain model based on understanding the structure of the brain.

This plan later received strong support and attention from the European Union, because the significance of this approach is not only to understand the intelligence of the human brain itself, but may even find alternative treatments for brain diseases.

But problems also come one after another. It is very difficult to simulate the entire human brain neural network using computers.

This is not only because of the complexity of computational simulation, but also because of the complexity of the biological brain itself.

After all, the human brain contains as many as 1011 neurons, and the amount of calculation and cost required are evident.

When humans actually use their brains to perform a series of actions such as reasoning and creation, they consume only 20-25 watts of power.

In other words, the biological brain has the characteristics of "high intelligence" and "low power consumption".

This is why studying biological brains is the best blueprint for general artificial intelligence.

And this kind of signal has begun to emerge.

For example, Single Cortical Neurons as Deep Artificial Neural published in the top journal NEURON in 2021

Networks research shows that-

A deep neural network requires 5 to 8 layers of interconnection Neuron can represent the complexity of a single biological neuron.

This also proves the powerful computing power of a single neuron. Therefore, if a single neuron can be characterized in a very fine manner, it can more closely approximate the complex process of biological information processing.

But the significance of simulating the biological brain in a more refined way goes beyond that.

Currently, humans still have many difficult diseases in the brain, such as Alzheimer's disease, depression, and brain damage.

The process of studying various brain diseases is a process that consumes huge manpower and material resources. If the biological brain can be accurately simulated, it may provide another possibility for solutions.

……

In short, to better simulate and understand the brain is to not only understand the brain itself, but also to value humans themselves.

The above is the detailed content of This nematode is not simple! The brain has been restored with high precision and can move forward dynamically. For more information, please follow other related articles on the PHP Chinese website!