In an article published in the journal Nature on june1,2022, researchers at the University of Pennsylvania introduced a powerful new optical chip developed by them The experimental device is composed of a neural network, which can process information in the form of light (it can process nearly 2billion images per second), without being delayed by traditional computer chip components such as memory.
Research diagram - 1: traditional optical electronic depth neural network (from: upenn)
Deep neural networks are the important foundation of this on-chip photonic chip for image classification. As a system modeled on the way the brain processes information, these networks are composed of interconnected nodes like neurons.
Research diagram - 2: implementation example of optical electronic neuron
Then, the training is carried out by feeding a large number of data sets, so that the neural network can learn the thinking mode similar to the biological brain - for example, it is used to distinguish objects in images or words in speech.
Research diagram - 3: structure example of photon classifier
With the passage of time, neural network will be better in related tasks. But the University of Pennsylvania study does not process information in the form of electrical signals - but in the form of light.
Research drawing - 4: image classification demonstration
By using optical fibers as their neurons and stacking them in multiple layers, each layer can focus on a specific type of classification work. During the test, the research team used a chip with a size of only 9.3m ㎡.
Extended Figure 1 - schematic of image formation and measurement settings
After the training of relevant data sets, the chip can classify the images containing two character sets, and the accuracy is as high as 93.8%. For the four types of image classification, the accuracy is 89.8%.
Extended Figure 2 - PDNN chip training and threshold calculation
The most impressive thing is that the chip can classify each character in 0.57 ns - which means it can process 1.75 billion images per second.
Extended Figure 3 - propagation time measurement
The research team said that this speed advantage over modern computer chips stems from the chip's ability to convert information into light for processing.
Extended Figure 4 - comparison with prior art
Firooz aflatouni, the first author of the paper, said that the chip processes information through the so-called "computation by propagation", and they skilfully skip the conversion steps of optical and electrical signals.
Extended Figure 5 - ways to enhance scalability and computing time
Different from the clock based system, propagation calculation occurs when light travels through the chip. This on-chip chip can directly read and process optical signals, making the processing speed of this technology much faster than in the past.
Extended Figure 6 - microring alignment algorithm and characterization
Another advantage of this technology is that there is no need to store the information being processed, so there is no need to send the data to memory - and the component is not needed at all - to save a lot of time and protect the data (prevent any form of potential disclosure).
Extended Figure 7 - block diagram of the electronic control circuit
Next, the University of Pennsylvania research team plans to expand the scale of the chip and make appropriate adjustments to the technology so that it can process other types of data.
Details of this study have been published in [nature] published on june1,2022( https://www.nature.com/articles/s41586-022-04714-0 ) ) On the journal, the original title was an on chip photo deep neural network for image classification.