Three Minutes Understanding the Core Technology of AI: Deep Learning

Since 2012, Deep Learning has broken through classical AI problems with a momentum of breaking through. Faced with the rapid development of artificial intelligence, don't you want to know its basic working principle?

To figure out what deep learning is, we need to start with artificial intelligence. Since computer scientists confirmed the term artificial intelligence at Dartmouth Conferences in 1956, there have been many fantastic ideas about artificial intelligence. We dream of having five senses (or even more) and reasoning abilities of human beings. And the magical machine of human thinking. Nowadays, although the dream situation has not yet appeared, slightly weaker artificial intelligence has become popular, such as image recognition, speech recognition, multilingual translation and so on.

Machine learning is an important method to realize artificial intelligence. The concept of machine learning comes from early AI researchers. In short, machine learning is to use algorithms to analyze data, learn from them and automatically summarize them into models, and finally use models to make inferences or predictions. Unlike traditional programming language development software, we use a large amount of data to send to machine learning, a process called "training".

Five hundred

Deep Learning is one of the most important parts of machine learning in recent years. Deep Learning is rooted in the neural network-like model. However, today's deep learning technology and its predecessors are quite different. At present, the best speech recognition and image recognition systems are all based on deep learning technology, as announced by mobile phone manufacturers. The AI photo function and Alpha Go in the past are all based on in-depth learning technology, just different application scenarios.

The foundation of in-depth learning is big data, and the path of implementation is cloud computing. As long as there is enough data and fast enough computing power, the "results" (presenting some kind of intelligent function of the machine macroscopically) will be more accurate. At present, the intelligent operation path based on big data and cloud computing can be better explained in the framework of deep neural network.

Deep neural network, also known as deep learning, is an important branch of artificial intelligence. At present, deep neural network is the basis of many modern AI applications. Since the breakthrough of deep neural network in speech and image recognition tasks, the number of applications using deep neural network has increased explosively.

At present, these deep neural network methods are widely used in automatic driving, speech recognition, image recognition, AI games and other fields. In many areas, deep neural networks are different from early experts in manual feature extraction or rule-making. The advantages of deep neural networks come from the ability to extract advanced features from raw data using statistical learning methods on a large number of data, thus effectively representing the input space.

In fact, the process of representation involves the process of computing a large amount of data, because the ultimate high accuracy for a particular function is at the expense of super high computational complexity.

Usually what we call computing engine, especially GPU, is the basis of deep neural network. Therefore, the method of improving the energy efficiency and throughput of deep neural network is very important for the wider application of deep neural network in AI system without sacrificing accuracy and increasing hardware cost.

Five hundred

At present, researchers of some well-known large companies in China have paid more attention to the development of special acceleration methods for deep nerve computing in recent years, and started to develop special chips for artificial intelligence, that is, real artificial intelligence chips.

The so-called AI chips are generally ASIC (special chip) designed by pointer to AI algorithm. Although the traditional CPU and GPU can also be used to implement artificial intelligence algorithms, these chips are either slow in calculation or high in power consumption, so many shortcomings make them impossible to use in many occasions.

For example, auto-driving vehicles need AI chips because they need to recognize the changes of pedestrians and traffic lights in the process of driving. These situations sometimes occur suddenly. If we use traditional CPU to do this sudden road condition calculation, because CPU is not full-time AI calculation, so it is necessary to use traditional CPU to do this sudden road condition calculation. Its calculation speed is slow, it is very likely that the green light has turned into a red light, and our autopilot has not been brakes.

If the GPU is used instead, the calculation speed will be much faster, but the calculation power consumption is very large at this time. The on-board batteries of electric vehicles can not support this function for a long time, and the high-power chip will cause the car body to heat up, which will easily lead to the spontaneous combustion of the fuel tank. Moreover, GPU is generally expensive, and ordinary consumers can hardly afford such an automatic driving vehicle that uses a large number of GPU chips. Therefore, in the field of artificial intelligence, the development of dedicated chips has become an inevitable trend.

Five hundred

(Implementation Path of Special Chip Technology Developed by Industry for Different Scenarios)

At present, AI chips available in the market can be divided into two categories according to the different processing tasks.

—— For training and inference, this work can be done by GPU, CPU and FPGA. But if we develop chips of artificial intelligence, we can do better. Because AI chips are professionals, they are equivalent to "experts".

—— Inference acceleration chip. This kind of chip is to put the trained model of the neural network on the chip to run. For example, the Cambrian Neural Network Chip and the DPU with a deep knowledge of science and technology.

Please read the Chinese version for details.