Photonics chip can train AI
Researchers at the University of Pennsylvania have developed a programmable photonic AI chip that can train neural networks in an energy-saving way.
Liang Feng and his team are researching energy-saving photonic chips. They have now been able to create an AI chip with a semiconductor that reacts to light, which they also switch with light in order to train a neural network.
(Image: Feng/University of Pennsylvania)
A team of researchers led by Liang Feng has presented a chip that relies on optics instead of transistor circuits and electrical currents and uses light particles as an information carrier and storage medium. Photonic chips are considered to be energy-efficient computers that could also perform AI in an energy-saving manner in the future. The special feature of the programmable photonic chip at the University of Pennsylvania is that it can map non-linear functions in addition to linear relationships. This ability is crucial for training AI applications. The underlying neural networks are characterized, among other things, by the fact that individual nodes only fire when their input exceeds the set threshold – a non-linear function.
The chip not only processes signal light, which is fed in as input and ultimately provides the output in a modified form. It is essential that the optical elements of the photonic chip consist of a semiconductor that reacts to light. These components can therefore be switched by light beams that are directed vertically onto the chip, so-called pump beams. The researchers used this function to create training algorithms for the implemented neural network: Depending on the output of the signal light and the difference to the desired behavior, the pump beams gradually varied and refined the individual parameters of the neural network.
Videos by heise
The chip itself is not changed by this adjustment, only the photonic properties of its optical elements, as Feng explains on the project page. This makes it the first programmable photonic chip for non-linear functions. In tests, the project team achieved a hit rate of 96 percent with its innovative AI chip after training when differentiating between different species in the so-called iris flower data set.
(agr)
