50 years of ups and downs in optical computing, new opportunities for AI catalytic photoelectric hybrid
Picture source: Nature
Author: Wu Wenhao
More and more artificial intelligence tasks require specialized accelerators to ensure fast and low-power execution of calculations. The optical computing system has thus entered the field of vision of more people.
Optical computing systems may be able to meet the needs of these specific fields, although after half a century of research, general optical computing systems have not yet become a mature and practical technology.
Artificial intelligence reasoning tasks, especially visual computing applications, are considered to provide opportunities for reasoning based on optics and photonics. From this perspective, Nature recently published a Perspective article Inference in artificial intelligence with deep optics and photonics , which reviews the application, prospects and challenges of photonic computing (including hardware and algorithms) in artificial intelligence over the past half a century. .
This article was co-authored by French, German, and Swiss researchers from Stanford University, Massachusetts Institute of Technology, and University of California, Los Angeles. The following is the key interpretation of the paper by the "data factualist".
Global Computing Power Contest
Computing power, an index that measures the computing power of computing equipment and algorithms, has now evolved from a concept explored in academia into an "arms race" among major powers on a global scale.
Whether it's autonomous driving, robot vision processing, smart home, or remote sensing technology, high-precision microscopy equipment, the Internet of Things, surveillance, and national defense, deep learning is widely used in these areas that contain astronomical data.
But to effectively run more and more complex advanced algorithms, the existing parallel computing power and bandwidth of GPUs and conventional accelerators (commonly referred to as AI chips) are not enough.
In the execution process of the algorithm, deep learning can be briefly divided into two stages, one is the training stage, and the other is the inference stage. There is a big difference between the two.
In the training phase, a deep neural network (DNN) requires a large number of labeled examples, and then for a specific task, an iterative method is used to optimize the parameters of the DNN. After the training is completed, the DNN can be used to perform inference. Currently, many technical solutions use GPUs for the inference stage of algorithms. However, due to the limitations of high power consumption, latency, and cost, it is impractical for many terminal devices that require deep learning algorithms to be equipped with GPUs, such as self-driving cars and IoT terminal devices.
In the face of such a scene that is difficult for GPUs to cover, the current solution is to use a more flexible AI accelerator. Broadly speaking, computing platforms designed based on optical principles can also be classified into this category. However, compared with traditional methods of performing calculations electronically, photonic computing platforms are more likely to have disruptive effects or even bring about disruptive effects in the AI field. Research paradigm shift.
The characteristics of light are inherently suitable for linear computing (the most important part of AI computing), which includes high-dimensional parallel computing. In the past few years, a series of optical chip start-ups have emerged and received investment from many technology giants with AI business. The main reason is this.
As the design based on optical principles is gradually being widely used in contemporary data centers, the industry's understanding of optoelectronics and integrated optics will gradually deepen, and it will be possible for information technology to enter the next era with the help of optical system design.
The history of photonic artificial intelligence
A list of the development schedule of optoelectronic computing related to AI, source of picture: Nature
This paper mainly discusses the development of photonic artificial intelligence in the past 50 years. However, the team traced the origins of photonic artificial intelligence to the time when Donald O. Hebb published The Organization of Behavior in 1949 . This neuropsychological theoretical work has been regarded as one of the most influential books in modern times, like Darwin’s "Origin of Species". The statistical learning theory and methods represented by deep learning are derived from neuroscientists and mathematics. The first proposed connectionism (connectionism), The Organization of Behavior is one of the foundation works.
From the 1960s to 1986, outstanding progress in the field of photonic artificial intelligence included: In 1957, the American psychologist Frank Rosenblatt proposed a neural network with a single-layer computing unit, the Perceptron; 1960 In 1964, Widrow & Hoff published "Adaptive switching circuits", they implemented neural networks with hardware, proposed ADALINE networks, and published Widrow-Hoff algorithm to propose adaptive switching circuits; in 1964, Vander Lugt proposed Optical Correlation laid the foundation for spatial light calculation; in 1982, John Hopfield proposed the Hopfield network, and Teuvo.Kohonen published "Self-organized feature maps", which introduced the SOM algorithm, which is a A simple and effective unsupervised learning algorithm...
After 1986, more AI advancements as we know them today were born. For example, in 1986, DERumelhart, Hinton, and others proposed the use of multi-layer perceptron with backpropagation, the important concept of backpropagation came out; in 1990, Yann LeCun and others used CNN to implement digital characters Recognition.
Among them, an important research work in 2017 was Chinese scientist Shen Yichen. This paper received by Nature Photonics demonstrated for the first time a programmable photonic processor for deep learning.
Not only computing platforms, but also optical algorithms continue to appear. There are optical CNNs in 2018, all-optical diffraction neural networks, and high-bandwidth photonic neural synaptic networks in 2019.
Advantages of optical computing
After reviewing the development history, the team started by building the basic elements of deep learning algorithms, and carefully analyzed why a design based on optical principles can have potential advantages in many aspects over the current existing design. In order to achieve these advantages, it may be necessary Technical problems overcome, and some specific and potential practical application scenarios (such as imaging and microscopy).
In short, optical computing systems use photons generated by diodes or lasers to replace electrons used to represent data streams in traditional designs to achieve energy efficiency superior to traditional computing device designs.
In the traditional design, the data information of 1 and 0 is represented by transistors. A single transistor can be used thousands or millions of times per second, but can only process one piece of information per use. The working efficiency of a transistor depends on the charging and discharging frequency of the wires connected to it and the heat released by this process. In order to represent and process more and more complex information, traditional designs need to add more transistors, causing the computer itself to run In the process, more heat will be released, which will cause problems in the design of the computer's energy consumption and volume requirements.
In contrast, in optics, since photons of different wavelengths do not affect each other, optical computers allow multiple information to be transmitted on the same information loop at the same time to achieve multiplexing (not only processing one information at a time) ). In terms of cycle time, optical transistors can cycle once every picosecond, which is a thousand times faster than electronic transistors.
In terms of energy consumption, the use of carefully selected wavelengths to transmit information can make the light not emit heat during the transmission process, reducing the energy consumption and volume requirements of the system, and photons, unlike electrons, do not undergo quantum tunnelling (quantum tunnelling, According to quantum mechanics, electrons can magically jump outside the potential energy wall that physically does not allow it to cross. This phenomenon stems from the fact that the probability of electrons running outside the potential energy wall is not zero)", so in theory, optical transistors can be compared Electronic transistors save more space, allowing us to make the transistor itself smaller.
However, the theoretical advantages of distance and the most important challenge facing the design of optical computing systems are also related to their design principles.
As mentioned earlier, the optical computing system can carry different information with different wavelengths, but if you want to process this information, you need to intervene in the information transmitted in the form of light, that is, use other electromagnetic wave signals to interact with it. The calculation process of the optical computing system itself is non-linear, and the strength of this interaction is much weaker than the electronic signal in the traditional design (the processing difficulty is greater than that of the electronic signal).
In addition, in theory, the cycle rate of optical transistors can be much faster than that of electronic transistors, but there is a theoretical conversion limit for electromagnetic waves in physics, which means that the information response speed of optical transistors will still be limited by the spectral bandwidth of the selected light, so , If you really want the optical computing system to have a higher computing power than the traditional electronic computing, it is necessary to have a practical method of transmitting ultra-short pulse signals with a high dispersion waveguide.
In the design of the computing system itself, optical logic gates need to be integrated into the high-level components of the computer (such as CPU), but these components are not optical components, so the optical computing system still involves converting optical signals into electrical signals. This conversion The process also affects the overall efficiency of the system.
The research team also said that at this stage, most of the optimism about the design of optical-based computing systems basically stems from the promise of ultra-low energy consumption for this type of design. Despite this hypothesis of ultra-low energy consumption, the efficiency of the photoelectric conversion process is often not taken into consideration.
Therefore, from the perspective of industrial applications, photons have not yet reached the dominant position of electrons in the entire technology industry.
One of the main reasons is that mankind's understanding and utilization of photons as bosons are far more mature than electrons as fermions. Especially in the overall industrialization process, electronics has gone through a complete route from the advent of electricity to the birth of electronic circuits, to product integration, engineering, and then industrialization, and photonics is far from completing these stages. For example, in high-tech industries such as optical communications, optical display, and optical storage, photonic devices are a key part, but at the level of various complete systems and equipment, compared with electronic devices, their proportion of total cost is The output value can be described as small.
At present, traditional electronic system design is still the mainstream of the industry, and most optical systems are for analog purposes.
The future of photoelectric hybrid system
Existing research on photoelectric computing in the field of imaging and microscopy (see Fig. 4. and Fig. 5 in the original paper for the specific paper corresponding to each small picture), source of the picture: Nature
In the specific AI application scenarios, the paper points out two important applications of existing optical computing: one is the field of macroscopic computational imaging. The optical computing system combined with deep learning algorithms can recover the scene from the results of single-pixel camera shooting. The original appearance realizes end-to-end optimization of optics and imaging processing.
The other is in the field of microscopy. Since we do not fully understand the process of light and matter, it is impossible to obtain microscopic images of many complex processes through modeling and analyzing data. We can only learn this unknown with the help of deep learning algorithms. The process of generating the "correct" image.
Schematic diagram of optical encoding + electronic decoding system, image source: Nature
In the conclusion, the authors stated that although the current pure photonic computing system design is still facing many practical technical problems in the specific implementation, many of these problems have been "hopefully solved".
In the past two decades, the design of computing systems based on optical principles has made great progress, mainly in the following aspects: nonlinear systems based entirely on optical principles, controllability of large-scale photonic computing systems, The photoelectric conversion efficiency and the programmability of the photoelectric conversion process are improved.
The design that combines traditional methods and optical methods is likely to be one of the most promising directions in the current optoelectronic computing field. The reason is that this hybrid design can combine the flexibility of the existing traditional design with the bandwidth and speed of the optical method, and retain the low energy consumption characteristics of the optical method to a certain extent, allowing future terminal devices to use deep learning on a large scale Algorithm becomes a reality.
Reference:
1、https://www.nature.com/articles/s41586-020-2973-6
About data combat faction
The data practitioner hopes to use real data and industry practical cases to help readers improve their business capabilities and build an interesting big data community.
