Scientists submitted a neural network operating at light speed

Researchers of the University of California presented a three-dimensional optical neural network, which allows for complex mathematical calculations at the speed of light.

A command using a 3D printer created a multi-layer neural network transmitting data by means of light pulses, which significantly increases the speed of information processing. Scientists say that instead of adjusting the calculations in each individual neuron, the optical network adjusts them by changing the phase and the amplitude of the beam. Therefore, the system is not attached to 0 and 1, and when solving neurons, either transmit or reflect photons.

During the test, the algorithm showed 55 thousand handwritten numbers from 0 to 9, training to recognize characters. Then offered 10 thousand images of numbers. As a result, the optical neural network at the speed of light correctly identified 91.75%. Although the efficiency is not yet high, but this is only the first sample and the team tested performance, and not accuracy.

Scientists expect their invention to help withdraw a new level of calculation underlying any artificial intelligence. Although the device cannot completely replace the graphics processor, but their symbiosis will significantly reduce the power consumption and heat generation of hardware, increasing the power.

Scientists from the University of Alberta promise to improve the efficiency of information storage devices,