Artificial intelligence seems to be surpassing humans in many areas, but one of the aspects where the human brain is superior is energy consumption. For a complex calculation, a human brain needs a power of 20 watts, while a supercomputer for a comparable task needs roughly a million times as much energy. That computer runs at 20 megawatts. Couldn’t that be done better, researchers have been wondering for a long time? Can’t we design hardware that is much more like the human biological brain?
One of them is Christian Nijhuis, who conducts research at TU Twente into so-called molecular switches. His team recently achieved a breakthrough, which was described in a trade journal Nature Materials. It’s high time to take a look at Nijhuis’ lab, where he works on the design and measurement of new molecules.
Nijhuis calls himself an ‘architect at the molecular level’. He draws molecules and if they seem interesting enough, chemists can make such a molecule. It’s called synthesizing. ‘We have now found a molecule that can mimic the function of a synapse,’ explains Nijhuis. The brain consists of about a hundred billion neurons, with each neuron connected to thousands of other neurons. The synapses are the communication officers: every signal in the brain goes through the synapses. Of these, there are an estimated one hundred to two hundred trillion, or a 1 with fourteen zeros.
For Nijhuis, the interesting thing about these synapses is not so much the unimaginable quantity, but their efficiency. That is why the synapse is the great source of inspiration when designing new computers. And that is badly needed: ‘The processing of the ever-increasing data streams is getting out of hand. Data centers are slurping energy and that’s only going to increase.’ Nijhuis mentions the well-known AI example: recognize the cat in the photo. AI is quite good at that these days, but it takes a huge amount of computing power to train models and then make the prediction ‘is this a cat?’ to be carried out.
Nijhuis: ‘Our brains do that much better. They only use energy when an information pulse passes through the synapse, which means they can process a lot of data at the same time.’ And not only that, they are also flexible and dynamic: ‘The synapses act as switches, but the special thing is that they can change all the time. So you want molecular hardware that also has those dynamic properties.’ So a computer that, just like a human brain, makes new connections every time it learns something new, strengthens or weakens existing ones, just what is needed.
Meticulous manufacturing and measuring process
This simulating of a single synapse has now been achieved. The first step was taken two years ago, with the arrival of a molecular on/off switch. And now there is the switch that also learns from the past. Nijhuis says with fire: ‘It was fantastic when we saw that our molecules behaved as we hoped.’
This jubilant moment was preceded by a meticulous manufacturing and measuring process. A very thin layer of exactly one molecule thick is applied to a gold plate, after which they are separated from each other again. Nijhuis elaborates on microchannels, alloys, mono- and oxide layers and stable electrodes. Monter: ‘That’s it, a very simple manufacturing method.’ In the lab, a student shows the slippery monolayer, with electrodes on both sides. He uses a multimeter to demonstrate that the molecular layer does indeed conduct.
The basement of the lab has been set up as a vibration-free Faraday cage, where the behavior of the molecules can be measured with special equipment. One molecule opens the doors to a whole family of new molecules and materials, Nijhuis hopes. And eventually to an artificial neural network in which the artificial synapses no longer process binary information flows (with zeros and ones), but work in an analogous way.
‘That is unprecedented’
It is still a long way off, admits Nijhuis. Johan Mentink, who conducts somewhat similar research at Radboud University, thinks so too. Mentink, who was not involved in Nijhuis’s research, calls the TU Twente’s discovery a scientific breakthrough: ‘They have succeeded in making a synapse on a molecular scale. That is unprecedented.’
Mentink himself is working on, among other things, optomagnetic synapses, which switch based on light. The big advantage of this is that they are much more economical. An important disadvantage is that they are not organic and are therefore less suitable for medical applications.
And it is precisely in this last direction that Nijhuis thinks with his organic invention, in addition to all applications where there is not an abundant amount of energy available, such as in self-driving cars or drones. Wouldn’t it be great if the body could use a super-powerful and economical computer made of soft materials instead of hard silicon chips, Nijhuis dreams aloud.
As an example, he mentions an implant that is placed on the visual cortex and can process the information coming from the eyes, as an aid for the blind. Careful experiments are already underway with implants, but these are coarse, hard chips with (in this specific example) limited results: the blind wearer sees an image of only a few tens of pixels.
Other useful, but less spectacular applications will come within reach sooner, Nijhuis hopes, such as devices that can administer insulin based on the eating pattern and behavior of a diabetic patient.
Don’t run straight to the health store now, warns Nijhuis: ‘Years and years of fundamental research will be needed before we are ready for practical applications.’