In terms of processing requirements, tasting wine can be quite a complex process.
As the human grape aficionado swirls the wine on their taste buds, neural networks fire up to process arrays of data.
Synapses gauge the importance of information such as sweetness and acidity before passing it on to the next layer as the brain parses the constituents of the flavour.
Learning to mirror the way that these natural neural networks operate is a major focus of artificial intelligence (AI) research today.
AI neural networks can match or even outperform biological brains in some areas, but they tend to need a lot of power to do so.
The human brain can process a wine tasting using as little as 20 watts of electrical energy, for example, while an artificial system may require thousands of times more.
Artificial neural networks may also be subject to hardware lag, generally making them much less efficient than our brains.
Researchers are constantly looking for ways to reduce the energy requirements, and a team from the National Institute of Standards and Technology (NIST) has developed a new kind of hardware that could do just that.
The new hardware used tech familiar from hard disk drives
Magnetic tunnel junctions (MTJs) are a type of device that can carry out some of the same calculations as a neural network using much less energy.
This technology has already been used for years in the read-write heads of hard disk drives and operates more quickly than conventional chips because it stores data in the same place as it carries out its computation.
What the scientists were unsure of was whether an array of MTJs could effectively function as a neural network.
The scientists, from NIST’s Hardware for AI programme, along with colleagues from the University of Maryland, were able to create a simple neural network using the devices and trained it to be a wine connoisseur.
The system was trained with 148 virtual wines from a full dataset of 178, with a range of factors to consider, including alcohol level, colour, flavonoids, and levels of ash and magnesium.
It was then let loose on the full collection, including the 30 wines it had not yet encountered, and was able to identify wines with a success rate of 95.3%.
The aim was not to create a wine bot, but to demonstrate the principle of using MTJs as neural networks, potentially leading to reduced-power applications in a number of different areas in the future.
Today’s news was brought to you by TD SYNNEX – the UK’s number one solutions distributor.
Read more of our latest Industry Updates news stories
