The new neural network operates at 300 000 km / s

Los Angeles (pte013 / 30.07.2018 / 12: 30) –

At the speed of light, a network of artificial neurons works by researchers at the University of California http://ucla.edu. Until now, such networks, which are modeled on the human brain, work with electric current, which moves much more slowly than light. It's "only" a speed of almost 300,000 kilometers per second.

Secrets of Photons

"Deep learning" refers to the ability of neural networks, such as the human brain, to learn continuously. Unlike humans, whose brain functions are large but ultimately limited, neural networks can outperform the brain. For example, deep learning has taught computers the complex Japanese Go game so perfectly that the best human players have no chance.

Neural networks based on photons can almost explode learning, believe Xing Lin researchers. They call their development "Diffractive Deep Neural Network". To make it happen, they created small plastic tiles using a 3D printer. These represent virtual neurons, that is, nerve cells, of which about 90 billion are in the human brain. Each artificial neuron has the same capabilities as its natural model. It can reflect or transmit incident light – in the human brain it is a weak current.

Currently still at an early stage

The photon-based neural network is still in its infancy. But the researchers were able to show that it works. This is a kind of feasibility study. They experimented with five plastic tiles between which there were small holes. In front of him was an object that had to be recognized. Then they fired the first tile with the laser light, which made its way to the other tiles. At the end, the photodiodes captured the light. From there, the object could be rebuilt.

Next, the researchers formed their network on numbers. Gradually, they added the numbers from zero to nine to the system. In this phase of learning, they used a conventional computer that works with electrons. Overall, they scanned the images of 55,000 numbers. The recognition value in the next test with the photon-based neural network was 95%. For example, according to researchers, a "full-fledged" system could be used to identify faces in crowds. Neural networks today do not do it with the necessary security.

(end)