How do neural processes work in the brain

"We want to simulate the function of the brain"

The human brain is made up of a hundred billion nerve cells, the neurons, that are linked by quadrillion synaptic connections. Even if scientists only partially understand this complex biological system, they can already describe some functions mathematically. In addition to computer models, researchers also use electronic circuits to simulate neural activity. In the journal “Science Advances”, Martin Ziegler and his colleagues from the University of Kiel report how the function of the brain can be simulated with so-called memristive components. In an interview with Welt der Physik, the physicist explains how they can even use them to technically simulate perception processes.

Martin Ziegler

World of Physics: What is a Neural Network?

Martin Ziegler: A neural network is the artificial replica of parts of our brain. Like a biological brain, it essentially consists of two components: the nerve cells, the so-called neurons, and the connections between them - the synapses. The bottom line is that the connections are not constant, that is, they change over time. This happens in the human brain when we learn to recognize a language or an object, for example.

How do neurons and synapses process information?

The job of neurons is to collect information by adding up the various incoming stimuli. The information corresponds to electrical voltages. If a certain potential threshold is exceeded, a signal is passed on in the form of small voltage pulses. Synapses then pass on these pulses, which are also known as spikes or action potentials. We would now like to recreate these synapses in neural networks, since memory processes in biological systems take place there.

Which methods can be used to implement this technically?

The majority of neural networks are implemented with the help of computer programs. These work with mathematical functions. In other words, the programs calculate what happens between the input and output of a neural system. We, on the other hand, work with so-called neuromorphic systems. We have developed new types of electronic components and connected them in electrical circuits to form a network.

What kind of components are they?

These are so-called memristive components. Memristor is an artificial word from the English words “memory” for memory and “resistor” for electrical resistance - so to speak, resistors with memory. These components consist of two metal electrodes and a so-called memristive layer in which ionic and electronic processes determine the charge transport. Applying spikes changes the resistance of this layer. A large part of our research is to create these elements. We don't want to recreate synapses or neurons as they occur in biology, but rather imitate biological systems on a functional and abstract level with the help of material physics and quantum mechanical effects. Simple circuits made of resistors and capacitors take over the function of the neurons and the memristive components take over the function of the synapses that are not constant over time.

Solid-state electronic component

What can be simulated with such electronic components?

We have implemented a circuit based on these solid-state electronic components that makes it possible to simulate the so-called binding problem. This is one of the big problems in neuroscience, which is how the brain can recognize objects. For example, there are various things that make a person stand out: their appearance, their voice, their smell or their characteristic gait. All of this information is processed in different areas in the brain. The question now is how the brain connects the information at the time of recognition and thus can recognize a person.

And how is that possible?

In the mathematical model of the physicist and neuroscientist Christoph von der Malsburg, the attachment problem is described with so-called neural oscillators and the synchronization of different areas. Recognizing a person on the basis of different sensations then means that the different areas of the brain are synchronized and active at the same time. If you know someone well, you don't need all the information to recognize them. It is therefore not necessary for all areas to receive information in order to be synchronously active at the time of detection.

Have you reproduced the mathematical model with your components?

Yes that's true. We had to pay particular attention to the fact that not the entire network was synchronized, but only different subnetworks. We can then change these different stable sub-networks depending on the context via the memristors. This can be made clear with the help of optical illusions: For example, if you draw a three-dimensional cube on a sheet of paper, there are two ways of perceiving the cube. Depending on which point you draw your attention to, the cube appears to protrude or protrude into the plane of the leaf. This corresponds to two different subnetworks in the neural network, between which we can switch back and forth.

Memristors

How can you switch back and forth between the subnetworks?

By changing the resistance of the memristive components - i.e. the synapses - we can influence the stability of the subnetworks. The input information of the neural network then corresponds to the “attention” and the output information is the perceived object, such as the drawn cube.

What applications could the neuromorphic networks have?

The idea is that new types of computers will evolve from the neuromorphic networks. Our computers today are very capable of arithmetic operations such as calculating the square root of a large number. Computers calculate this much faster and better than humans. However, people are significantly better than the most powerful computers at cognitive tasks such as pattern recognition. With the neuromorphic systems we are now trying to develop a new type of data processing that comes much closer to how the human brain works.