What neuron modeling software are you using

"We recreated a neuron"

Computers are fast computers - they solve billions of arithmetic operations within a second. But when it comes to versatility and efficiency, the human brain is still by far the best thinking machine. It controls our movements effortlessly and requires just as much energy as a 20 watt incandescent lamp. Researchers like Gianaurelio Cuniberti from the Technical University of Dresden want to find out how such an achievement can be transferred to new technologies for computers. In an interview with Welt der Physik, the physicist talks about slime-covered computer chips and how they can learn.

World of Physics: You are looking for new concepts for the computers of tomorrow. What bothers you about them today?

Gianaurelio Cuniberti

Gianaurelio Cuniberti: Computers can now calculate a lot very quickly. But compared to human and animal intelligence, they cannot make decisions. That is why researchers developed artificial intelligence. However, it is still a pure software thing to this day. We write programs that process information like our brains. But these programs need an incredible amount of resources - that is, powerful computers. We wanted to take a different approach and develop an artificial object that really works like a brain. This saves us the software and on top of that, the energy consumption is significantly lower.

So our brains are inherently highly efficient computers?

Exactly. On average, our brain needs as much energy as a 20 watt light bulb. It doesn't matter whether we're the new Einstein or a completely normal, average person. However, a computer that mimics the function of the human brain would require as much energy as a nuclear power plant provides. Apart from that, neither the hardware nor the software for such a task has yet existed. However, there has already been great progress. For example, today we would probably lose a game of chess against a computer. But when it comes to chess, the rules are clear.

And these are not always recognizable at first glance in "real" life?

Correct. When it comes to everyday things, the brain is still far better and, above all, more efficient than all artificial machines. And not just ours, but also that of primates and other mammals. So the question is where does this efficiency come from? Researchers have found that nerve cells work in the human brain. We call them neurons. Each of these neurons is connected to an incredible number of neighbors. That's what we call connectivity. In modern computers, the central computing element is called the Central Processing Unit, or CPU for short. Many small cells, the transistors, also work in it. But each of these transistors only communicates with a very few neighbors. In addition, the CPU is only responsible for processing the data. They are saved in a different location, namely in the main memory or on the hard drive. In all of our computers and smartphones, processing and storage take place in separate locations. In the human brain, on the other hand, there is no such separation. Information is stored and processed in the same place.

Are there other factors that make our brain so efficient?

Yes, because the brain is changing. For example, when we practice juggling a ball, the nerve pathways involved in that movement are trained. Over time, they pass the electrical impulses on faster and faster. We are becoming more skilled. Our brain succeeds so efficiently because it is not only very well networked and trainable, but also stores and processes the data in the same place.

If nature has perfected the brain and nerve cells over millions of years, why not take advantage of this and build a computer out of nerve cells right away?

The idea is not that far-fetched. Scientists have tried to use the capabilities of nerve cells in the past. In the 1960s, for example, there were tests in Germany in which two electrodes measured the impulses of individual nerve cells in the laboratory. Squids are particularly good for this. They have long nerve cells that can even be seen with the naked eye - this is ideal for these experiments. But if you were to build a chip on this basis, the cells would have to stay alive and be supplied with nutrients. In addition, they only survive when it is extremely wet around them. Such a computer would be anything but practical.

In the laboratory

Then how did you manage to recreate how a brain works?

We used electronic components instead of living cells. For example, we artificially recreated a neuron in the laboratory for the first time. While conventional transistors have poor connectivity, i.e. are connected to a few neighbors, our neurotransistors - as we have called them - have an incredible number of neighbors. After we stimulated them with certain information, we were able to show that they would later pass the electric current on more quickly when they received similar information. So you've learned. We also managed to get the neurotransistors to store information.

How did you do that?

Our transistors can have several neighbors and also work differently than their relatives in smartphones or computers. You can think of a conventional transistor as a channel through which current flows. For example, if you press a button, this triggers an electrical impulse that opens or closes the channel. All modern computers are built on this principle. Our approach is different. We let the stream itself open the channel. If more current flows, the pressure in front of the channel increases and it opens without a command. This is exactly what happens in the brain when we learn something.

And how do you get conventional electronic materials like silicon to behave like this?

The secret is Solgel. It's a viscous, almost slimy plastic. We applied it as a very thin layer on silicon platelets. There it hardens and becomes porous. Ions, i.e. charged particles, then move between the many pores. However, they move very slowly and it is precisely this inertia that we can use and store information for a certain time, namely as long as it takes the ions to move back. The information is saved directly where it was processed. And the more a single transistor is excited, the faster it opens the channel and lets current flow. This strengthens the corresponding connection - as in the brain. The system learns.

So your neurotransistor mimics the way the brain works and is less like a conventional computer chip. How could such chips be used in the future?

Computers based on our chip would be less precise and would estimate mathematical calculations rather than calculate to the last decimal place. But they would be smarter. A robot with such processors would, for example, learn to walk or grasp with them. He could have an optical system with which he can not only register images, but also recognize connections in what is seen. And all of this without having to develop a single line of software. We call this neuromorphic computer. And not only can they be built much smaller than conventional processors, they are also much more energy-efficient. This means that they can be installed directly at the data source - for example in earth observation satellites, in the head of a robot or in its legs. And they can adapt to changed tasks during operation and optimize them independently.