On a Friday afternoon, Ton Juny Pina tries to imitate the smelling brain of a fruit fly. On the table in front of him is a clear plastic box containing a battery of gas sensors that look like small thimbles. The researcher at the University of Groningen can flow different gas mixtures through the box, to which the sensors all respond. “We have sixteen sensors that are sensitive to different gases, but to be honest, such sensors are not very selective: they typically all respond somewhat to a very wide range of gases,” says Juny Pina.
This makes designing an artificial nose, a sensitive and specific detector for odors, no easy task. That is why he and his supervisor Elisabetta Chicca feed the signals to a computer model of the tiny brain Drosophila melanogastera fruit fly that uses less than a million brain cells to navigate flawlessly to every organic bin or abandoned apple slice.
“We use a model of the corpora pedunculatathese are brain circuits that insects use to separate the signals from different odors,” says Chicca, professor at the Groningen institute CogniGron, focused on neuromorphic computingwhere operating principles of the brains, nerves, senses of people and animals serve as inspiration.
“The ideas behind neuromorphic computing have been around since the 1980s, but things have been moving very quickly in the last ten years,” says Beatriz Noheda, physicist and until recently director of CogniGron in the new Feringa research building on the Groningen campus. The center, founded in 2016 with the help of 40 million euros from the estate of an anonymous donor, employs more than a hundred researchers from various disciplines, from psychologists and neuroscientists to computer scientists, physicists and materials scientists.
The lab of Elisabetta Chicca and Ton Juny Pina.
Photos Willem Serne
But there are also almost a hundred research leaders at twelve other universities, companies and research institutes who are now conducting research into this, according to a white paper published this year by the Top Sector ICT. The Netherlands is well positioned to become a neuromorphic world power, the report argues.
Applications are in areas where fast and energy-efficient data processing is needed, such as monitoring electricity networks, logistics, security, disaster relief, food production and healthcare, as well as in the development of energy-efficient computer chips and sensors. Academic research has already led to spin-offs, such as the companies Innatera, Hoursec and AxeleraAI, which develop energy-efficient chips or sensors.
This rise is in various ways related to that other technology boom inspired by biological brains, known to a wide audience as AI. Virtually all chatbots, image generators and other amazing and terrifying feats that we now have AI and machine learning are based on deep neural networks. And its operation is in turn inspired by neurons, the brain cells that now help you understand this text.
Although both computer AIs and real brains are much more complex than can be explained in the next paragraph, the basics are simple: neurons or brain cells, real or simulated, transmit signals to each other. Each neuron adds up the incoming signals and determines whether the total exceeds a certain limit. If so, that neuron itself also sends out a signal to the neurons connected to it further down the network.
The weight, the amplification factor with which the signals are transmitted, can change per connection. By adjusting these weights, the network as a whole can be trained: the network of neurons learns to process information, whether it is recognizing someone’s voice or generating a deepfake.
Deep learningnetworks consume enormous amounts of energy. ChatGPT is really becoming a problem
But despite the enormous success of artificial neural networks, there is also something to be said: in addition to their sometimes disappointing, mendacious and soulless intellectual performance, they must be trained with immense amounts of data. And „deep learningnetworks consume enormous amounts of energy. ChatGPT is really becoming a problem,” says Noheda of CogniGron. “While our brains run on about 20 watts, and can do even more.”
The neuromorphic inspiration goes beyond the idea of neural networks, simulated by digital computers, says Noheda. “Neuromorphic computing can mean many different things: it can actually refer to any principle of the brain.”
One of those principles has to do with a fundamental difference between brains and computer chips, says Yoeri van de Burgt, a researcher at TU Eindhoven, one of the Dutch pioneers in the area. “Digital computers work according to the Von Neumann principle: you do a calculation in a processor, and then you send the result to the memory. For the next calculation, you retrieve data from the memory, and so you are constantly funneling data back and forth.”
When simulating large neural networks, you quickly run into the ‘Von Neumann bottleneck’: everything has to happen one after the other, while in the real brain it runs parallel. Today’s AI boom only started when researchers discovered that graphics computer chips from Nvidia, intended for rendering 3D computer games, bypass this bottleneck. Each chip contains multiple processors that perform calculations in parallel. This allows you to quickly perform large ‘matrix multiplications’, an operation that involves multiplying and adding large tables of numbers. Matrix multiplications are common in 3D simulations, but calculating the signals between neurons in a neural network also amounts to a matrix multiplication.
“Parallel processors speed up computation, the next step is to completely eliminate the separation of memory and computation,” says Noheda. In the brain, the calculation units (the neurons) and the memory units (the weights) are also evenly distributed over the gray matter. “One approach that is currently being strongly pursued is to also do this on a computer chip,” says Noheda.
A memristor from TU Eindhoven.
Photo Pei Zhang
Such a neuromorphic chip consists of a grid of horizontal and vertical lines, which are connected to memristors at the intersections. The word is a combination of memory and resistor. This is an electrical resistor that can vary in value, and which can also remember the value for a long time, like a local memory. The whole can calculate a matrix multiplication very quickly without the need for processors or transistors, and thus also a calculation step in a neural network. The Groningen spin-off IMChip is working on the development of a neuromorphic computer chip.
This approach works best in combination with an older neuromorphic idea: that of spiking networks. Van de Burgt: “One of the reasons that the brain is so efficient is that it works with short pulses of electrical activity. Neurons receive such spikes in series: the shorter the time between the pulses, the stronger the signal.”
By adding up the pulses that arrive within a certain time, the neuron decides whether it should also fire pulses itself. “So the time between successive pulses contains the information, and that is very efficient. The spikes themselves require very little energy, and most of the time nothing happens.”
In the 1980s, computer chip pioneer Carver Mead and his PhD students at the California Institute of Technology were already working on cameras, hearing aids and other sensors that imitated biological senses. The pixels in such a camera only fire their pulses when the value of that pixel changes. So if nothing changes on the screen, no data is sent. Such event-based cameras, with fast response but extremely low energy and data consumption, are now on the market. Portable heart monitors or other medical monitors that measure biological signals, and which preferably require a long battery life, also work according to this principle.
“Mead was inspired by physicist Richard Feynman’s famous motto: I can’t understand what I can’t make,” says Noheda. The people who obtained their PhDs at Mead, and later the researchers who worked with them, still call themselves neuromorphs,” says Noheda, “and our Elisabetta Chicca, from the artificial fruit fly nose, is one of them.”
Brains have evolved over millions of years. To survive we usually don’t need great precision
While neuromorphs like Chicca were initially mainly neuroscientists, computer scientists and electronics scientists, in the past ten years or so physicists and materials scientists have also been included in the field, such as Noheda himself.
New neuromorphic chips are no longer based on transistors and CMOS technology (complementary metal oxide semiconductor) with which they are etched in silicon, but on other combinations of materials that behave like pulsating neurons and the aforementioned memristors. “For this you need something that changes resistance, but in a predictable and controllable way,” says Noheda. “A system, for example, is a network of silver nanoparticles that touch each other here and there. When more current flows through them, they melt together a little and become better conductors.” Noheda itself is working on memristors based on ferroelectric materials, a variant of the ferromagnet that you encounter on refrigerators.
At the Amsterdam physics institute Amolf, Bruno Ehrler’s research group discovered an extremely energy-efficient memristor, based on the material perovskite, known from super-efficient solar panels, and Yoeri van de Burgt’s research group in Eindhoven is working on a memristor based on electrically conductive polymers.
“Unlike silicon, it is a soft material,” says Van de Burgt in his laboratory in Eindhoven. On the lab table lies a tiny piece of transparent plastic, connected to three electrodes. “It is conductive to electric current but also to hydrogen ions. By pumping them in with an electrode, the electrical resistance changes.”
Such a memristor cannot become very small or fast, “but the idea is not to cram millions of them onto a chip to compete with Nvidia,” says Van de Burgt, who is more likely to think of medical applications. “This material, in combination with other special polymers that deform when heated, could potentially serve as an artificial muscle.” The group has already experimented with the use of neurotransmitters, signaling substances in the brain, to control the polymers. Van de Burgt: “And because it is soft, it is better suited to medical applications, such as heart monitors, or human-friendly soft robots.”
For the time being, we can continue to learn and copy biological brains, says Noheda. “Brains have evolved over millions of years. To survive we usually don’t need great precision, but we do need energy efficiency and quick decisions. So that’s what the brain is good at.”
The journalistic principles of NRC