System could overcome problems with artificial intelligence on ‘edge’ devices
An “unprecedented” new computer chip could help revolutionise AI, according to its creators.
The system should allow increasingly complex artificial intelligence to live on chips themselves, without having to send information to the cloud.
At the moment, most applications of AI are not done locally on a device, because they have limited battery lives and processing power. Instead, information is sent over the internet to another computer, where it is analysed or computed and then sent back down again.
Eventually, experts hope that artificial intelligence will be able to be embedded in “edge” devices: objects such as phones that could perform detailed AI tasks whenever and wherever.
The new breakthrough is a step towards that, since it allows for a wide variety of different AI tasks to be done far more quickly and efficiently than ever before.
Usually, such efficiency is seen as being available at the cost of versatility, and chips can either use less power or do more tasks, but not both. The new system appears to overcome that problem, however.
It does so using “resistive random-access memory”. That allows for computing to be done directly in the memory, rather than shifted into separate processing units, which speeds up the processing time.
The system has already proven itself incredible capable, both in its usage of energy as well as the amount of time each one takes.
It showed 99 per cent accuracy in analysing handwritten digits, for instance, and 84.7 per cent on a Google speech recognition task.
But scientists hope to improve it, making it even quicker, more efficient and ready for a wider spread of uses.
A paper describing the findings, ‘A compute-in-memory chip based on resistive random-access memory’, is published today in Nature.