When a wine expert tastes a new wine, the neural network in his brain process a collection of data from each sip. Like the human brain, scientists want an artificial intelligence-based (AI) system to be data connoisseurs (taste experts). For this, they have developed a computer version of neural networks to analyze and process data, TechXplore reported.
Image Credits: Pixabay
But an AI system consumes more energy than a human brain to perform tasks. Also, the hardware can lag and becomes less efficient. Due to this, scientists in the AI research field are looking for alternatives that can consume less energy.
Researchers at the National Institute of Standards and Technology (NIST), U.S., were able to design new hardware that could be more efficient and consume less energy. It has also passed the virtual wine-tasting test.
The research team suggested that magnetic tunnel junction (MTJ) could be alternative hardware to build AI’s neural network. MTJ is a device consisting of two magnetic metal layers such as cobalt-iron separated by a thin insulating layer of aluminium oxide which helps in electron conduction. This device is good at analyzing neural network uses and requires less energy to work. Also, the other devices based on MTJs consume less energy than their traditional hardware counterparts. MTJs work more efficiently and quickly because they store data and perform calculations in the same place. These are commercially important as they serve as information processors for hard disk drives.
Though researchers were confident about the energy efficiency of MTJs, they wanted to check whether the collection of these devices could work as a neural network or not. For this, researchers took MTJs for a virtual wine-tasting test.
Along with the NIST’s hardware for the AI program, scientists with their colleagues from the University of Maryland, U.S., program a simple MTJ neural network provided by their companions at Western Digital’s Research Centre, California.
To work as a better wine expert, the AI system needed to improve its virtual palate (virtual intelligence). The team used 148 wines from a dataset of 178 made from three types of grapes to train the network. Each virtual wine had 13 characteristics to consider, such as ash, magnesium, alkalinity, colour, alcohol level and flavonoids. Each characteristic was assigned a value between 0 and 1 so that the network could distinguish between different types of wines.
The network then tested for virtual wine-tasting on the full dataset which included 30 wines. It passed the test with a success rate of 95.3 percent. Out of the 30 wines the system hadn’t tested on, it made two mistakes. This showed that the network was working properly.
The researchers concluded that the array of MTJ devices could be used to develop new AI systems. They said that using these devices as synapses (connections between two neurons) could reduce energy consumption while processing data.
“It’s likely that significant energy savings over conventional software-based approaches will be realized by implementing large neural networks using this type of array,” said Jabez McClelland, NIST Physicist.