Artificial intelligence produces remarkable results but uses too much energy. In the future, synapses made from a solar cell material and cast into hardware could help.
The spectacular “works of art” achieved with artificial intelligence (AI) have been making headlines in virtually all media for years now. In contrast, intelligent processing processes are less well-known to the public but are increasingly proving to be a success factor in virtually every sector of the economy. What is more, this is possible even though AI models can replicate the biological structure of our brain only in a highly simplified manner. Unfortunately, they use huge amounts of energy. While human thought organs get by with 20 watts, a deep learning model developed by scientists at the University of Massachusetts releases nearly 300,000 tons of CO₂ into the air.
The main reason for this is the fact that with conventional computers data need to be moved between the processor and memory all the time. In contrast, nerve cells in the brain take care of both data storage and data processing.
In order to apply the much more efficient principle of neurobiological circuits to technical systems, scientists have been carrying out research into so-call memristors (a portmanteau of memory and resistor) for many years now. The electrical resistance of these dual-pole, passive components is not constant—it depends on what quantity of electrical charge has flowed in which direction. This value is retained even without a power supply.
Various operating modes exist for memristors. Depending on the architecture of the artificial neural network it is advantageous to use all these modes. To date however, memristors could only be configured for one such mode in advance. Researchers at the ETH Zürich, the University of Zürich and Empa have now developed a concept which allows a switch between two operating modes “on the fly”. With the first mode, the signal is reduced over time and eventually disappears (volatile mode). With the second, the signal remains constant all the time (non-volatile mode).
The human brain works in a similar manner. Biochemical messenger materials on the synapses transmit stimuli between the nerve cells. The stimuli are initially strong and then slowly become weaker. And during learning, the brain forms new, longer-lasting synaptic connections with other nerve cells.
The “Swiss” memristors consist of halide perovskite nanocrystals, a semiconductor material which is used in photovoltaics. “Stimulus conduction” takes place as a result of silver ions from an electrode arranging themselves temporarily or permanently to form a nanofiber which the perovskite structure penetrates, thus allowing electricity to flow.
This process can be influenced so that the silver ion fiber is either thin and gradually breaks down again into individual silver ions (volatile mode) or thick and permanent (non-volatile mode). Low currents result in volatile mode, and strong currents in non-volatile mode.
According to the researchers, it will be possible to produce computer chips with memristors which support both modes. After all, it is generally not possible to combine different memristor types on a single chip.
However, improvements need to be made before memristors can replace “energy guzzling” neural networks. Until then, they can still make an important contribution towards neuroscience research. After all, they are quite similar to real neurons.