How super computers are inspired by the human brain

• "Neuromorphism" is a way to go beyond the limits of computers by taking inspiration from the architecture of the brain, according to our partner The Conversation.

• This process replaces lengthy calculations on a "classic" processor with an ultra-fast physical phenomenon that can be integrated into conventional electronic circuits.

• This analysis was conducted by Chloé Chopin, postdoctoral researcher in neuromorphic spintronics, and Flavio Abreu Araujo, professor of neuromorphic spintronics.

Our computers are getting more and more powerful.

They are gaining in speed, in computing power, and make it possible to carry out ever more complex tasks, as evidenced by the recent switch to exascale supercomputers with "Frontier" in the United States.

These advances are crucial to continue to progress in cutting-edge areas of computing such as artificial intelligence and to be able to deal with problems of increasing complexity.

Although it is possible to increase the number of transistors to increase the computing power or the memory of a computer by miniaturizing them - up to a certain point (a process for manufacturing transistors as small as 2 nanometers is currently in development), this is not enough.

Indeed, there is another limit which is inherent in the organization of the components of a computer.

This limit comes from the fact that the processor which performs the calculations and the memory which stores the results are physically separated.

The computer then devotes more time and energy to transferring data between memory and the processor than to performing the mathematical operations necessary to perform useful tasks.

This phenomenon is called the “von Neumann bottleneck”.

One way to overcome these limitations is to rethink the architecture of a computer and come up with new components that are inspired by the human brain.

This new type of brain-inspired electronics is called “neuromorphism”.

Neuromorphic devices have already proven themselves, for example, by recognizing written numbers as well as spoken numbers and vowels.

​Using “spintronics” to mimic the functioning of neurons

These neuromorphic devices can be manufactured in particular thanks to spintronics, a form of electronics which exploits both the charge of the electron (like conventional electronics), but also its “spin”.

Thanks to spintronics, long and energy-intensive calculations on a "classic" processor are replaced by an ultra-fast, energy-efficient physical phenomenon that can be integrated into conventional electronic circuits.

We can also create new strategies to bring memory units closer to computing units and thus obtain more powerful, faster and more energy-efficient computers.

Of course, there are still many challenges to overcome, such as connecting a large number of structures together (to build complex circuits) or developing learning algorithms optimized for this type of circuit with a neuromorphic architecture and spintronic components.

Spintronics, késako?

The spin of the electron is a quantum property that is difficult to represent in the classical world.

A simplified vision is to imagine an electron as a small magnet which turns on itself.

This small magnet can be oriented up or down and when a current has more electrons with an upward spin (or the reverse), the current is said to be “spin polarized”.

Two very important phenomena in spintronics are sources of applications: “magoresistance effects” and “spin transfer”.

“Magnoresistance effects” are exploited in computer hard drives.

The principle is as follows: two magnetic layers are separated by a non-magnetic material (which is either conductive in the case of “giant magnetoresistance” or insulating in the case of “tunnel magnetoresistance”).

One of the magnetic layers has a fixed magnetization, like a fridge magnet, while the other has its magnetization which can move, this is the “free” layer.

If the two layers have parallel magnetizations then the polarized current flows easily and the resistance of the structure is low while if they are antiparallel, then the current flows with difficulty and the resistance of the structure is high.

“Spin transfer” arises from the reverse effect.

In this case, it is the polarized current which will transfer its spin moment to the magnetization of the “free” layer with the effect of reversing the latter.

In some cases, this spin transfer phenomenon also leads to sustained oscillations of the magnetization to create what are called spin transfer oscillators.

​Making artificial synapses and neurons

By combining both spin transfer, to manipulate a magnetic layer (writing), and the magnetoresistance effect, to measure the resulting resistance state (reading), it becomes possible to imagine innovative devices that will be able to draw inspiration from, or even imitate, certain characteristics of the human brain in order to manufacture synapses and artificial neurons.

A very large number of synapses connect the neurons to each other, transmit information and play the role of small memories while the neurons integrate the information received and send back a signal called "action potential" if a threshold is exceeded, such as a small computing unit.

During learning, synaptic connections will either strengthen or weaken.

They will thus transmit more or less information according to their importance.

It is therefore interesting to manufacture structures whose resistance can vary thanks to spin transfer in order to produce artificial synapses.

In practice, several approaches are explored by researchers

To manufacture artificial neurons using spintronics, several approaches are possible.

One of them is to seek to reproduce what is observed in the brain, by using more or less complex architectures to manufacture artificial neurons, in order in particular to reproduce threshold effects, or to generate trains of "

” (action potentials) like a biological neuron.

A second approach is to create an artificial neuron whose response is "non-linear" (which is a fundamental property for artificial intelligence).

For this, it is possible to use spin transfer oscillators like those described above.

The “free” layer can also have a particular magnetic structure called “vortex”.

It is the most stable state for dimensions ranging from 50 nm to 5 microns.

Its transient dynamics give it short-term memory in addition to nonlinear behavior.

Our "BRAIN" file

This structure is promising for neuromorphic applications and it is studied in particular to be able to model its behavior.

Indeed, by developing an accurate and rapid model, one can, to a certain extent, dispense with experimentation and simulations and save considerable time: what would potentially take months or even years of simulation is reduced to a few minutes of calculations.

This work is part of a promising way to study complex architectures based on spintronic components and thus overcome certain limitations of conventional computers.

high tech

IA: Will machines be able to demonstrate musical creativity without human assistance?

high tech

Why tomorrow's artificial intelligence will be (much) faster… and greener

This analysis was written by

Access to this content has been blocked to respect your choice of consent

By clicking on "

", you accept the deposit of cookies by external services and will thus have access to the content of our partners

I ACCEPT

• Science

• Video

• The Conversation

• Brain

• Computer

• Computer science