Science and technology development In 1956, the concept of "artificial intelligence" was proposed. Since then, after more than sixty years of ups and downs, artificial intelligence has experienced two waves in total.
The second wave is based on the "if-then" structure, which is based on the artificially set formal logic; the second wave is more common nowadays, based on big data driving, with the help of statistical methods, simulated neural networks, etc. to realize the computer Independent learning.
From the transformation of these two waves, we found that the essence of AI is computing, and algorithms control the flow of data and achieve so-called "intelligence." Computing is undoubtedly an important part in the development of artificial intelligence. Now, where is the calculation of artificial intelligence?
Light speed calculation, solve the cost problem of AI training
Recently, OpenAI has investigated and analyzed the amount of calculation consumed by large-scale AI experiments in different periods. It was found that the amount of calculation required for AI training has increased by 30 times compared with 6 years ago, which is equivalent to doubling every 3.5 months.
The increasing amount of calculation is actually not a bad thing, because this means that the ability of AI is also increasing day by day, but the improvement of computing power also makes the training cost of AI continue to increase. Take the AlphaGo Zero that everyone knows now as an example. This is the current large-scale AI experiment, and its cost may be 10 million US dollars. If the amount of trial calculation continues to increase, its cost will increase by an order of magnitude every 1.1-1.4 years. According to this trend, in 5-6 years, the cost of this experiment will reach 20 million US dollars.
Unless there are some very powerful AI technologies that can bring large-scale economic returns, otherwise, to maintain the AI ​​computing trend and ensure that the next "Alpha Dog" can be "fed" out, the economic output must be To grow by orders of magnitude. This is just the current situation. No one can guarantee that the trend of AI computing will not rise faster in the future.
Therefore, the current problem facing enterprises and governments is how to speed up the speed of AI operations to meet the increasing amount of computation in the research of artificial intelligence.
Speaking of speed, the current fast speed in the universe is the speed of light, and the speed of light propagation is 300,000 kilometers per second. If the signals in the deep neural network of AI can be propagated at the speed of light, can the speed of calculation be increased accordingly?
Recently, researchers from the University of California, Los Angeles (UCLA) used 3D printing technology to print out a solid-state neural network, and used hierarchically propagated light diffraction to perform calculations and achieved the effect of image recognition of handwritten digits.
The use of light to perform calculations is actually a natural fit with one of the classic algorithms of machine learning-linear regression algorithms. Linear regression generally estimates actual values ​​based on continuous variables, and the amplitude and phase of light are adjustable. Variable, this is also the difference between AI light speed calculation and electric field propagation in traditional computer circuits. The development of this technology is believed to be able to contribute to the continuous reduction of AI computing costs and mass production goals.
What does it mean to perform calculations with light?
With the landing of technology, what will happen to artificial intelligence when light speed computing technology is truly applied? Yan Xuan, an analyst of the Intelligent Theory of Relativity (ID: aixdlun), believes that there will be major breakthroughs in the following two aspects.
1. "Black box" has become a "glass box" under the sun
If human beings are the "God" of AI, then humans give AI only the rules of the combination of "life", and the real evolution is done by AI itself.
In the initial stage, AI's cognition is very limited, it seeks good results through continuous trial and error, and a certain degree of intelligence emerges. From a human perspective, we are absent in the process of AI cognition. Under the framework of deep learning, we "know it but don't know why," which is the "black box" problem.
AI predicts that you will die in 50 years, but you do n’t know how it works; the driverless car hits the guardrail next to the road, and you do n’t know what the problem is, you can only send it back to the original factory to revise Algorithm.
Let the AI ​​have the light, the problem of "black box" may be solved. You should know that although AI operations are invisible and intangible numbers, the diffraction of light is a real physical phenomenon. If the prediction process of the model is solidified into a physical representation, the process of artificial intelligence operations can be clearly observed.
In experiments that used light to perform calculations, UCLA researchers developed a 3D printing AI analysis system. This system can analyze artificial intelligence through the diffraction of light. The researchers also said that by changing the phase and amplitude, each "neuron" in artificial intelligence will be adjustable.
2. AI "cultivation" game: the opening of strong artificial intelligence
Can strong artificial intelligence (intelligent machines with consciousness and self-awareness that can reason and solve problems) be realized?
Some people predict that in the 21st century, there will be AI comparable to human intelligence. The position of this prediction comes from a computing subject based on reductionism. Its basic view is that the physical world, life processes, and even the human mind are all algorithmically computable.
The human brain works like a computer. As long as we can simulate the calculation rules of the human brain, we can build an intelligent machine at least equivalent to the human level. Of course, there is a hidden assumption that all human consciousness is the product of brain calculations.
How to create a strong artificial intelligence, we may start from the calculation of the human brain. A natural assumption here is that if we create an AI, we can have enough computing power from the age of zero to simulate the human brain running for 18 years, and capture the brain's intellectual performance with a fine enough granularity. This AI Can the problem be solved like an 18-year-old adult?
And how much is this calculation? There are many calculations for the peak speed per second (FLOPS, also known as "floating point operations per second") required to simulate the brain for one second. For example, the data collected by AI Impact results in a median of 1018 FLOPS and a range of 3 Between × 10 ^ 13FLOPS and 1 × 10 ^ 25FLOPS. Running such a simulation for 18 years is equivalent to 7 million petaflops.
With the continuous development of AI technology, its granularity will only be more subtle and the amount of calculation will be greater. If you can use light to perform calculations, it will undoubtedly provide a feasible technology for this imagination.
The "hard wound" of speed of light computing
The use of light to perform calculations has certainly revolutionized the calculation method of neural networks, but this method itself still has some problems.
First of all, in the experiment mentioned above, the operation of light is based on the solidified neural network. Therefore, when deep learning has completed training and all the values ​​of the parameters have been determined, it is then carried out using 3D printing technology. After curing, the printed neural network can no longer be programmed.
Secondly, it is very difficult to build an ultra-high-precision diffraction plate that can perform on-demand processing tasks. When solving the problem of calculating training costs, it is difficult to say that this new technology will not bring about the cost of hardware research and development. In addition to the manufacturing process, there are also difficulties in hardware installation and environmental stability.
It is true that the application of new technologies still needs a period of time. Whether it is enough to perform calculations to meet the high-speed growth of AI computing trends still needs our active exploration.
With the solution of computing problems, artificial intelligence will also make great progress.
Zhuzhou Zhirong Advanced Material Co., Ltd , https://www.zrcarbide.com