A breakthrough at the University of Twente brings new brain-like computers one step closer. An international group of researchers led by Prof. Dr. Christian Nijhuis has developed a new type of molecular switch that can learn from past behavior. The researchers published their findings today in the scientific journal Nature Materials. Nijhuis: “These molecules learn in the same way as our brain.”
Computers, data centers and other electronics use enormous amounts of energy. We are now building huge wind farms to meet that energy demand. But according to Prof. Dr. Christian Nijhuis, we can also focus our attention on making our electronics more efficient. “Our brains are the most efficient computers we know of. They use ten thousand times less energy than the most efficient computers,” says Nijhuis.
This is because our brains process data in a completely different way. Where computers process binary streams of information – with zeros and ones – our brains work analogously using time-dependent impulses. “Our brains process information from millions of nerve cells from all our senses without any problems. Unlike traditional electronics, they only use the brain cells and synapses through which the pulses pass,” says Nijhuis. Because energy is only consumed during one pulse, our brains can process a lot of data at the same time much more efficiently.
Hardware
The molecules Nijhuis and his team developed can perform all the Boolean logic gate circuits required to deep learning. “Deep learning is a form of machine learning based on artificial neural networks and is widely used for automatic recognition of images and speech, but also in the search for new medicine and more recently for making art. All things that are much more difficult for a computer than for our brain,” says Nijhuis. Scientists are making great strides in artificial intelligence software, but these molecules are now also bringing artificial intelligence hardware closer.
To mimic the dynamic behavior of synapses at the molecular level, the researchers combined fast electron transfer with slow proton coupling limited by diffusion. This is similar to the fast pulses and slow uptake of neurotransmitters by the neurons in your brain. The molecules can adjust the strength and duration of the pulses. Thereby they show a form of classical conditioning. The molecules adapt their behavior to the stimuli they have previously received. A form of learning. In the future, such molecules may also respond to other stimuli such as light.
This breakthrough makes it possible to develop a whole new series of adaptive and reconfigurable systems. These in turn can lead to new multifunctional adaptive systems that greatly simplify artificial neural networks. Nijhuis: “This drastically reduces the energy consumption of our electronics.” Multifunctional molecules that are also light-sensitive or can detect other molecules can lead to new types of neural networks or sensors.
Christian Nijhuis leads the group ‘Hybrid Materials for Opto-Electronics’ (HMOE; Faculty of Applied Sciences), part of UT’s MESA+ Institute for Nanotechnology. He is also principal researcher of the research area Computing Molecules & (Opto)Electronics within the Molecules Center in MESA+. This research was carried out in collaboration with Damien Thompson, Professor of Molecular Modeling and Director of the SSPC (Science Foundation Ireland Research Center for Pharmaceuticals at the University of Limerick) Enrique del Barco, Pegsus Professor at the University of Central Florida.
The release with the title ‘Dynamic molecular switches with hysteretic negative differential conductance emulating synaptic behavior’ was published in the scientific journal Nature Materials. Nature Materials is a top-3 journal in chemistry, physics and materials science. The publication can be read online.