Alex from Aofeisi Quantum Position | WeChat QbitAI
What is faster than deep neural networks?
Perhaps photonic DNN can answer this question.
Currently, a photonic neural network (photonic deep neural network, PDNN) developed by researchers in the United States allows image recognition in just 1 nanosecond.
What does 1 nanosecond mean? It equals 10-9 seconds, which is comparable to the single clock cycle of the most advanced microchips (the smallest unit of time).
Additionally, researchers found that the PDNN achieved accuracies of 93.8% and 89.8% for binary and four-class image classification, respectively.
Indeed, today’s large multilayer neural networks are efficient and powerful, but they are also limited by hardware and often require significant power resources.
However, the PDNN developed by engineers at the University of Pennsylvania can analyze images directly, without the need for clocks, sensors, or large storage modules, effectively reducing energy consumption.
This research was published in a paper that appeared in Nature magazine on June 1.
Photonics DNN is Faster Than Traditional DNN
How do the principles and performance of photonic DNN differ from traditional DNN?
First, let’s look at the traditional DNN:
Figure a shows the structural diagram of a traditional DNN, which includes a data arrangement unit, followed by an input layer, several hidden layers, and an output layer that provides classification output.
Figure b displays the structure of traditional N-input neurons: the linear weighted sum of inputs, passed through a nonlinear activation function, generates the output of the neuron.
Figures c and d are schematic diagrams of a PDNN chip’s neural network and N-input neuron structure, respectively.
First, the input image is formed on a 5×6 optical raster coupler, then arranged into four overlapping sub-images, whose pixels are sent to the first layer of neurons, forming a convolution layer.
The subsequent neurons are fully connected to their previous layer, producing two outputs that can classify up to four types of image information.
For these neurons, their inputs are optical signals.
In each neuron, linear computations are performed optically, while nonlinear activation functions are implemented optoelectronically, allowing classification times to be under 570ps(=0.57ns).
The paper’s corresponding author, electrical engineer Firooz Aflatouni, further described the performance of this PDNN: it can classify nearly 1.8 billion images per second, whereas traditional video frame rates range from 24 to 120 frames per second.
The PDNN chip circuitry is integrated within an area of just 9.3 mm2, without the need for clocks, sensors, or large storage modules.
A laser is coupled into the chip to provide light sources for each neuron; the chip contains two 5×6 optical raster couplers, serving as input pixel arrays and calibration arrays, respectively.
However, the uniformly distributed light supply provides each neuron with the same output range, which evidently allows for scaling up to larger PDNNs.
Image Classification Tests of Photonics DNN Chip
The researchers had this PDNN microchip recognize handwritten letters.
A set of experiments tested the binary classification performance of the PDNN chip: it needed to classify a dataset composed of a total of 216 letters “p” and “d”.
The chip achieved an accuracy of over 93.8%. ((92.8%+94.9%)/2)
Another set of experiments tested the four-class classification performance of the PDNN chip: it needed to classify a dataset composed of a total of 432 letters “p”, “d”, “a”, and “t”.
The chip’s classification accuracy was over 89.8%.
These results indicate that even with more classes (as in the four-class case), and variations and noise caused by printers, the PDNN chip still achieved high classification accuracy.
To compare the image classification accuracy of this PDNN with traditional DNN, researchers also tested a DNN consisting of 190 neurons implemented using the Keras library in Python, which showed a classification accuracy of 96% on the same images.
Author Biography
The first author, Farshid Ashtiani, is currently a postdoctoral researcher in the Department of Electrical and Systems Engineering at the University of Pennsylvania, focusing on photonic-electronic integrated systems.
The other authors of the paper are also from the Department of Electrical and Systems Engineering at Penn.
Last year, a scientist from Japan’s NTT Research Institute indicated that photonic computing could reduce the energy consumption of neural network computations, holding great potential and likely becoming a key focus of development in deep learning’s future.
The Penn engineers involved in this research stated that the PDNN’s direct, clockless processing of optical data eliminates the need for analog-to-digital conversion and large memory modules, making next-generation deep learning systems’ neural networks faster and more energy-efficient.
What are your thoughts on the prospects and applications of photonic deep neural networks?
Paper link:https://www.nature.com/articles/s41586-022-04714-0#article-info Reference link:https://spectrum.ieee.org/photonic-neural-network
— End —
Live Registration | Array Optical Waveguide:
Display Technology to Bring AR Glasses to Consumer Market in Three Years
The development of the metaverse industry is bringing together various cutting-edge technologies to build a new form of the next-generation internet. AR devices may become the “gateway” devices of the next-generation internet.
What is the core competitiveness of consumer-grade AR devices? What is the technical principle of AR optical modules? What is the current state of their development? What will your first AR device look like? On June 16, the live broadcast will reveal all~
Click here👇 to follow me, and remember to star it～
One-click triple connection “Share”, “Like” and “View”
Technological advances meet every day~