Neural networks are sweeping the world of computing. With their help, researchers are able to advance the process of machine learning. Face recognition, object recognition, natural language processing, machine translation ... These were originally human-only skills, and have now become routine machines.
As the neural network can promote the development of artificial intelligence, which gives researchers more incentive to create more powerful neural networks. The key to this research is to create circuits that resemble neurons, the neuromorphic chip. So, how to make the circuit speed is significantly improved?
Now, perhaps the answer to this question. According to MIT, the Alexander Tait team at Princeton University created the world's first photoelectron neural network and demonstrated its computational speed.
Optical calculations have long been high hopes. Photons have a higher bandwidth than electrons, so you can process large amounts of data faster. However, optical processing systems are not widely used because of their high cost. In the analog signal and other tasks, this ultra-fast data processing capability only photon chip can provide.
Neural networks now offer a new opportunity for photonics. "With the help of the silicon photonic platform, the high-speed information processing capabilities of photonic neural networks can be used in areas such as radio, control computing," said Alexander Tait.
At the heart of this photonic neural network is an optical device. Each of its nodes has neuron-like response features. These nodes are in the form of miniature circular waveguides that are etched into a silicon base that allows light to circulate. Once light is input, it modulates the output of the laser operating at the threshold. In this region, small changes in the incident light will have a significant impact on the laser output.
Each node in the system uses a certain wavelength of light, a technique known as wave division multiplexing. Light from each node is fed into the laser, and the laser output is fed back to the node, creating a feedback circuit with non-linear features. This output is mathematically equivalent to a device called Continuous Time Recurrent Neural Network (CTRNN).
The Tait team said the device can greatly expand programming techniques for larger silicon photonic neural networks.
Researchers use a network of 49 photon nodes to simulate neural networks and how photonic neural networks can be used to solve mathematical problems in differential equations.
Tait compared it to a normal CPU. "In this task, the effective hardware acceleration factor of photon neural networks is about 1960 ×," Tait said. "This is a 3-order speed."
The researchers said the study opens the door to a whole new industry of photonics. Tait said: "Silicon photonic neural networks may be the first to enter the broader category of silicon photonics systems for scalable information processing."
Lamp Box Fabric Coating,Glow In The Dark Paint,Glow In The Dark Spray Paint,Glow Paint
Guangzhou Quanxu Technology Co Ltd , https://www.silicamattingagent.com