New microchip is faster than AI, claims MIT
Tue 27 Jun 2017
Researchers at MIT have created a silicon microchip which they believe to be faster and more efficient than AI.
The Institution has designed the photonic chip to improve the development of optical neural networks that will function in a similar manner to the way an artificial neural network would work.
The findings published in the journal, Nature Photonics, revealed that deep learning neural networks now have the advantage of understanding different behavioural patterns through light and sound.
Prior to this discovery, photonic processing was considered too time-consuming and impractical, but the millimetres-in-length chip could now help to store a greater amount of information than ever before.
The chip, comprised of a network of 16 neurons over four layers, functions when data enters via a beam of light energy and is split into four separate beams. Each light represents a unique number composed of information, which produces a different number depending on the brightness it produces upon its exit.
The paths of light passing through the chip cross and interact with each other, amplifying and weakening their individual intensities depending on the strength of the beam.
Optical computing using photons was a key factor in the AI study as, once the light beams have been generated through the chip, they are capable of working independently without guidance.
The study found that the light beam energy processed information more effectively than electricity could. Using sound vowels from recordings of 90 people, the researchers found that the chip had a 92% success rate in identifying different vowels, compared to standard neural networks that scored 77% in their experiments.
This significant boost now has the potential of helping AI develop at a much faster rate, with potential applications in a wide variety of cases from data centres to autonomous cars.