Be like a person. Machines for people, people for machines

We can give each other a lot. Or in fact we have already given. This is evidenced by countless biomimetic technical solutions that imitate processes in machines associated with human organisms and many other living beings. On the other hand, the story of how machines support us is so obvious that it seems trivial.

The most ambitious part biomimetism there are attempts to recreate in the inanimate world. The brain does not need a central processing unit (CPU) or a hard drive. It works by opening and closing nerve membranes and sending out waves of charged ions. These waves cause changes in the nerve endings that allow the brain to function, a process called synaptic plasticity that allows us to learn and process information.

Hypothetical artificial computing constructs that mimic the human brain are called "neuromorphic" - this term was created in the 80s by an American scientist Carver Mead. He was known for his attempts to mimic neurons with specially configured transistor circuits. Scientists around the world have long been hard at work on systems that work like the human brain. For this reason, they are sometimes called artificial brains.

What can he do that even the most advanced artificial intelligence system still cannot cope with? For example, it can quickly gather a lot of chaotic information into meaningful conclusions. That's why building a computer that can process and record information simultaneously—like a brain that instantly analyzes and stores large amounts of data—is one of the greatest technological challenges today.

Like real neurons

All traditional computers and computer-like devices are based on the architecture developed by John von Neumannin which the processor and memory systems are separated from each other. This is a handy solution when we want the machine to run in different programs. However, the processors, when they are running, are reaching for data from memory, and the transfer continues. A computational unit waiting for data often remains idle.

The volume of data from databases created by humans and machines (operating within the Internet of Things) is currently enormous. The von Neumann architecture becomes the bottleneck in these processes. The US Department of Energy estimates that computer data processing currently consumes between 5 and 15% of the world's energy production, including quite a bit for the transmission itself. It can be improved over the years quantum and photon methodsthat require less energy, faster and more efficiently. However, their use still remains a promising direction. Experts are paying more and more attention to the solutions offered by the human brain.

NS16e motherboard with sixteen IBM TrueNorth processors

Attempts to create flexible and effective learners neural networks undertaken since the 50s. The pinnacle of achievements in this area is Google's DeepMind, which in 2016 defeated the grandmaster of the game "Go". However, this synthetic brain is a software simulation of a neural network, and physical calculations are still performed in silicon systems. So the bottleneck of von Neumann is not overcome here.

Physically, the neural network mimics an IBM neuromorphic processor called TrueNorthbuilt in 2014. The problem is that if he were to use his architecture to create a completely synthetic equivalent of a brain, such a computer would need 10. times more energy than what works in the human head ... In addition, TrueNorth, although it can recognize simple images, does not work as connections between neurons and synapses. He is not able to strengthen certain connections in the learning process.

Intel has been working on neuromorphic processor technology for several years to mimic the way the human brain works. In the end, he boasted of the first device of this type. Experimental chipset called Loihi in 2018, he will move to leading universities and research institutes dedicated to the development of artificial intelligence. The company writes: "Intel has invented a first-of-its-kind, self-learning neuromorphic chip, codenamed Loihi, that mimics the way the human brain works by acting on the principle of stimuli of varying strengths received from the environment." The processor consists of 130 thousand. silicon "neurons" connected by 130 million "synapses". According to Intel, it is a thousand times more energy efficient than conventional learning machines.

Avoid silicone

TrueNorth chips or Intel's solution still stick to silicon, which has its advantages, but in terms of power efficiency, it doesn't match the organic matter of nerve cells. Therefore, scientists are working on completely different models of neuromorphic computers that will mimic the plasticity of the brain and allow processors to mimic synaptic function, for example, using a phase transition in certain materials that melt and solidify at certain temperatures, while maintaining the ability to store and release large amounts of energy.

Last summer, a group of scientists led by Evangelos Eleftheriou from the IBM laboratory in Zurich reported in Nature Nanotechnology that she was able to build artificial version of a neuron. It consists of a layer of germanium-antimony telluride between the electrodes. Here, the phase change of this material is used, which, depending on the applied voltage, passes from an insulator to a semiconductor, and then to a conductor. According to the researchers, this mimics fluctuations in the behavior of neurons. There are other projects in which phase changes in artificial synapses are affected by a light wave, which means significantly less energy consumption. So this is not the only suggested solution.

Researchers at the University of Southampton demonstrated in late 2016 that memristors and resistors are able to remember their previous resistance values ​​and can be used to build complex and advanced neural networks. As part of their experiments, they prepared a network of metal oxide-based memristors and used them as artificial synapses.

Illustration of connections between neuron and dendrite

As a result, they initiated learning process without outside interference – just as it happens in the human brain. The memristors created by scientists not only consume less energy than previous solutions, but also remember their previous state. These are electronic components that limit or regulate the flow of electric current in a circuit and are able to remember the amount of charge that has passed through them and retain data even when the power is turned off. Essentially, they perform a function similar to synapses and have the intrinsic ability to perform computational tasks and store information at the same time with much less volume and power loss.

A research team from the University of Southampton has developed memristor integrating sensor (Memristive Integrative Sensor - MIS) at the nanoscale, into which he introduced a series of "voltage-time" patterns that reproduce the electrical activity of nerve cells. It is reported that MIS sensors based on metal oxides, which function like the synapses of brain cells, are able to encode and compress (up to 200 times) the activity of nerve cells recorded using multielectrode arrays. The researchers said that in addition to bandwidth limitations, their approach is extremely energy efficient, as the power required for each recording channel was XNUMX times lower than it is currently.

However, the task of neuromorphic constructors seems to be many times more difficult in the light of new discoveries by scientists from the University of California at Los Angeles - it turns out that the brain has more than a hundred times more processing power than previously thought. According to these findings, dendrites, once thought to be simple passive signaling channels, turned out to be very electrically active, generating ten times more impulses than the soma (nerve cell body).

he remarked Mayank Mehta, a UCLA neurophysicist describing the California study in the media.

Invisible, visible thanks to machines

The simulation of our brain is not very good for machines yet. In return, however, we received "from them" - and still receive - that which allows us not only, for example, to fly into the sky or lift weights, but also to see and hear what our senses cannot catch. We have had night vision devices, thermal imagers, ultraviolet detectors and sonars for a long time. These devices give us access to abilities previously only seen in animals or insects. Relating to mechanisms from the natural world, they remain products of technical thought.

A typical example of modern technology that opens our eyes to new worlds is lenses with an ultra-thin graphene detector operating in the full infrared range. They were created at the American Michigan University as a result of research by the group of prof. Zhaohui Zhong. Another group of scientists and engineers, led by Joseph Ford from UC San Diego and Erica Tremblay from the Institute of Microengineering in Lausanne - in turn, she developed contact lenses with a polarizing filter, similar to those worn in 3D cinemas, allowing you to see with almost three times the magnification.

Of course there are also augmented realitywhich not only allows doctors to look inside the human body without surgical intervention, but can also help, for example, firefighters quickly navigate and search for people in fire conditions when visibility becomes poor or even zero. C Thru Helmet has a built-in thermal imaging camera, the image from which is transmitted to the fireman in a helmet directly on the display in front of his eyes. On the other side Stryker II system, created by BAE Systems for aircraft pilots, integrated with a helmet, equipped with sensors that automatically set the pilot's glasses to night mode with night vision.

You can finally reach cameras that see the invisible. Even things around the corner of the building. The invention that made this possible came from scientists at the Universities of Bonn, Germany, and British Columbia, Canada. It is based on the reproduction of images outside the field of view using diffused light. This method uses a laser beam projected onto a wall, obscuring what should be observed through the camera lens. The device collects many different light reflections and sums them up, trying to create the resulting, i.e. image outline. Initially, you just see the wall in the camera's viewfinder. However, after a while, when subtle mathematical algorithms begin to work, revealing the so-called image echo (i.e., a small amount of light reflected from an object, diffused and then falling on the surface of the wall), we begin to see figures hidden around the corner. No living being can do such things anymore!

There are also methods for "See" something in the darkand they don't need more photons per pixel. For example, the one that developed Ahmed Kirmaniego from the Massachusetts Institute of Technology (MIT) and published in the journal Science. The device, which he and his team developed, Kirmani emits a low-power laser pulse in the dark, which, when reflected from an object, writes a single pixel to the detector. The principle itself is not new. New is a sophisticated algorithm that requires far fewer photons to create an image than before. It has been calculated that only one hundredth of what is required for currently used light detectors in difficult conditions, such as in LIDAR technology, is sufficient.

Looking at a running engine in AR technique

Since there are augmented reality-based opportunities to "see" the interior of a running car engine with the help of sounds coming from there, it may be time a camera that "reads" the sounds from the image? The prototype of such a device was developed by MIT, Microsoft and Adobe. The corresponding algorithm recorded speech sounds coming from a bag of potato chips, from which the recorder was separated by soundproof glass. In other experiments, it was possible to reproduce sound from a silent video. This method is based on "inference" about sounds by analyzing the movement and vibration of objects.

It's hard not to notice that the drive towards biomimetism and neuromorphism, as well as further solutions that strengthen our senses, clearly show how modern technologies, despite so many fears, are still focused on people.