❐ Definition and Scope of Artificial Intelligence
Artificial Intelligence (AI：Artificial Intelligence) means the intelligence exhibited by the machines manufactured by human being. The research scope of AI is very wide, including deduction, inference and problem solving, knowledge representation and reasoning, planning and learning, natural language processing, machine cognition, machine social, creativity, etc. We have frequently heard about “Machine learning” being a part of AI; furthermore, “Deep learning” belongs to a part of machine learning, as shown in Fig. 1.
Figure 1: Scopes of Artificial Intelligence, Machine Learning and Deep Learning.
❐ Classification of Artificial Intelligence
Artificial Intelligence may be classified into five levels as below depending on the capabilities the machine (computer) can handle and judge:
➤ First level AI: First level AI means the machine (computer), such as vacuum cleaner and air conditioner, has an automatic control function that can automatically make the corresponding response through a control program based on the signals, like ambient temperature, humidity, brightness, vibration, distance, image, voice, etc., detected by sensors. Actually, this is only a computer having an automatic control program and the programmer must consider all possible situations to write the control programs, so it is not an actual “intelligence.”
First level AI is just like an intern in a company, who can only execute the instructions from the boss and perform various repetitive jobs, but cannot consider if the instruction is right or wrong. For example, the boss asked someone to move large boxes to an area labeled as “large” and small boxes to an area labeled as “small” then the intern will only follow the boss’s instruction.
➤ Second level AI: Second level AI means the machine (computer) can explore the deduction and apply knowledge. Second level AI is a typical artificial intelligence, which utilizes algorithms to generate correlation between input and output data for creating extremely large amount of arrangements and combinations for input and output data. The possible applications include: puzzle analysis program, medical diagnosis program, etc.
Second level AI works like an employee in a company, who can appreciate the rules given by the boss and make decisions. For example, the boss wants to classify different sizes of boxes based on length, width and height and apply knowledge to identify different commodity types, such as fragile, inflammable. Then, the employee must measure the dimensions of the boxes according to the instruction and make classification, and judge which commodity to be fragile or inflammable.
➤ Third level AI: Third level AI means the machine (computer) can learn how to generate correlations between input and output data according to the data. “Machine learning” means the machine can learn the rules itself according to the input data. The possible applications include: search engine, big data analysis, etc.
Third level AI is just like the manager in a company, who can learn the rules and make decisions by himself. For example, the boss indicates the judging rules (features) for large box and small box that the manager should learn how large a box to be a large box. The manager can think by himself how large a box to be large according to his past experience.
➤ Fourth level AI: Fourth level AI means the machine (computer) can learn by itself and appreciate the “features” used to represent data during machine learning, so called “feature learning.” The possible applications include: Google taught the computer about the features of a cat.
Fourth level AI is just like the general manager in a company, who can discover the rules and make decisions. For example, he might discover a box which is rather big but is round (feature), so it should be processed differently from other commodities.
Third level (mainly for machine learning) and Fourth level (mainly for deep learning) are not easily differentiated. In fact, deep learning was developed from machine learning, and the major difference is that, during data processing of Third level AI, the “features” must be supplied by human for the machine (computer); during data processing of Fourth level AI, the “features” may be learned by the machine (computer) itself. This is a great breakthrough in AI field, which will be described in detail in the following articles.
❐ History of Artificial Intelligence
Since the humans invented the first computer, the associated development for artificial intelligence has started. The development has lasted for over half century and has gone through three upsurges, but there was no breakthrough each time due to certain technology difficulties. We will first introduce the development history of artificial intelligence and the reasons and encountered difficulties for each upsurge.
➤ First upsurge (1950~1960): ”Program search and deduction” started to develop in 1950s, which utilized the computer to perform searching and deduction for specific problems and solving them. However, the operation capability of the computer at that time was limited, and there would be at its wits’ end once coming across complicated problems. Thus, it was called “artificial intelligence” only for solving toy problems, and the upsurge was cooled down during 1960s.
➤ Second upsurge (1980~1990)：”Knowledge input and judgment” started to develop in 1980s, which inputted a large amount of expert knowledge into computers, then the computer could determine the answers of user’s questions. The expert system was applied in disease diagnosis. However, if there were one wrong judgment in a series of questions, a wrong result would be derived. Besides, knowledge should be infinite, so it is impossible to input all knowledge into computers and to find out the sequence of all knowledge. Thus, this technology became impractical and the upsurge cooled down during 1990s.
➤ Third upsurge (2000~)：”Machine learning and deep learning” started to develop in 2000s. Due to the progress of semiconductor technology, the computer gained great computation power. The cost of semiconductors was lowered so the cloud storage became cheap. So-called “Big data” was collected in the cloud server from across the entire world that established an excellent development foundation for AI. In which, Machine learning means using Big data to train computers “learning” the features of data, and Deep learning means using Big data to train computers “appreciating” by itself the features of data, which was also called “feature representation learning.”
Owing to the progress of semiconductor technology and cost down, the storage and computation for a large amount of data became easier which provides an excellent development environment for AI. According to the current development of the entire technology industry, we don’t have to worry about the development of AI slowing down someday; on the opposite, we might have to worry about the excessive development of AI would bring us some negative influence in the future.
❐ Cloud side and Edge side of Artificial Intelligence
Internet is an opened space formed by billions of computer hosts and servers connected together, as shown in Fig. 2(a). How do we describe the system constituted by so many computer hosts and servers? Scientists used a term “Cloud side” to represent such a system, with respect to “Edge side” where the users are located, as shown in Fig. 2(b).
Figure 2: Cloud side and Edge side of Artificial Intelligence.
The large amount of learning and computation for AI are currently performed in powerful processors in cloud servers. The Central Processing Unit (CPU) developed by Intel was employed in early stage. Then, scientists discovered the Graphics Processing Unit (GPU) developed by Nvidia, exhibits the performance about more than 100 times of CPU. Intel acquired the technology of Field Programmable Gate Array (FPGA) by merging Altera to compete with GPU. There are many more manufacturers starting development of Application Specific Integrated Circuit (ASIC), for example, Tensor Processing Unit (TPU) designed by Google or Vision Processing Unit (VPU) designed by Intel, which are both integrated circuits developed specifically for AI. These processors are all mounted at Cloud side. However, not all the applications are suitable for transmitting Big data to be processed at Cloud. For example, the self-driving car must process the data in the car (Edge side) to be able to reflect the road conditions in real time.
Apple Inc. recently announced iPhone X using the A11 processor developed by themselves. The embedded dual-core Neutral Engine (NE) is dedicated to process machine learning, deduction model and algorithm associated with image recognition, which is also an integrated circuit developed specifically for AI. The difference is that the A11 processor is mounted at Edge side, that is the user’s cell phone, so the cell phone may “automatically learn” recognizing the facial features of the user. Apple Inc. emphasized that all the facial features of the user are recognized and completed at the cell phone side and will not upload to the Cloud for processing, so there will absolutely be no doubt for data leakage. iPhone X, provided by Apple Inc., can make the user truly sense On-device AI at Edge side this time. In the predictable future, how the processor at Edge side combine with AI to become “Edge intelligence” will definitely be a hot topic.
【Remark】The aforementioned contents have been appropriately simplified to be suitable for reading by the public, which might be slightly differentiated from the current industry situation. If you are the expert in this field and would like to give your opinions, please contact the writer. If you have any industrial and technical issues, please join the community for further discussion.