New optical chip enables sub-nanosecond image classification

Text |

The human brain has billions of neurons, and there are thousands of connections between each neuron and other neurons. Many studies hope to create artificial neural networks to imitate the human brain to improve information processing capabilities. In recent years, with the development of science and technology, the development of deep neural network has achieved a qualitative leap. Deep neural networks provide a lot of support in areas such as computer vision, speech recognition, medical diagnosis, and more.

Currently, consumer-grade image classification technology on digital chips can perform billions of calculations per second at speeds sufficient for most applications. But more complex images, such as recognizing moving objects, 3D objects or human cells, still face obstacles. In addition, in the field of optics, despite good progress in photonic computing, the lack of scalable optical chips and the loss of photonic devices limit the advancement of optical deep networks.

Researchers at the University of Pennsylvania have developed a powerful new optical chip, the first deep neural network chip to be implemented in a scalable manner entirely on an integrated photonic device, measuring just 9.3 millimeters (0.01 inches). , which can classify nearly 2 billion images per second.

Traditional electronic devices typically use digital clock-based platforms such as graphics processing units (GPUs) for data processing, where computational speed is limited by clock frequency and memory access time. Also, raw image data often needs to be converted into digital electronic signals, which is not only time-consuming, but also requires large memory units to store images and videos, which can raise potential privacy concerns.

Unlike traditional electronic systems, the chip processes information in the form of light, using optical cables as neurons, stacked in layers, each dedicated to a specific type of classification. In the test, the chip can achieve an accuracy of 93.8% for the classification of image sets containing two types of characters, and an accuracy of 89.8% for the recognition of four types of character sets.

In addition, the chip is able to classify each character in 0.57 nanoseconds, allowing it to process 1.75 billion images per second. At the same time, it does not need to store the information being processed, which greatly saves time and ensures data security.

At present, relevant research results have been published in the journal "Nature"; in the future, the new chip can promote the development of face recognition, autonomous driving and other fields.