Lei Feng network press: Author Li Chao, go out to ask nlp engineers. The article is his insight after participating in the CCF-GAIR of Lei Feng Net in August, some conclusions on artificial intelligence, and the thinking of how individuals/enterprises can make artificial intelligence products. In August, I was fortunate to receive an invitation from Lei Fengwang (searching for the "Lei Feng Net" public number) to participate in the CCF-GAIR conference. The splendid guest group and the grand meeting were not to mention. As an engineer who has been working on nlp technology and related products, write personal ideas and gains here. It is divided into three parts: the first part is what the artificial intelligence is good at, and what is not good at it; the second part introduces the participation of the big gods in the field of machine learning, vision, etc.; the third part describes the thinking of how individuals are doing artificial intelligence products. . Professor Michael Wooldridge, head of the Department of Computer Science at Oxford University and head of the Deep Mind-Oxford Partnership, gave a conference report titled "Routes to Artificial Intelligence." This speech drew the author's long thoughts and summed up the following thoughts. Artificial Intelligence, abbreviated AI, was proposed at the Dartmouth Conference in 1955. John McCarthy defined it as "scientific and engineering of making intelligent machines." After the collapse of the AI ​​bubble in the 70s of last century, this concept has been quiet for some time. Related researchers and engineers are not willing to say that they are researching artificial intelligence, but have become machine learning, data mining, natural language processing, speech recognition, image recognition, etc. each field. Machine learning is a method, data mining, natural language understanding, speech recognition, and image recognition are goals and specific applications. So far, the progress of AI is mainly the progress of machine learning. In recent years, the improvement of computing power and the accumulation of Internet big data, deep learning has been applied in the fields of speech recognition, image processing, etc., and has made great breakthroughs. The concept of artificial intelligence was once again concerned by academics, industry, investors, and even the general public. Especially when AlphaGo defeated Li Shishi this year, it pushed AI to the highest point in history. People in various specific fields have returned to the embrace of artificial intelligence. To be cheeky to say that he is an artificial intelligence practitioner. Without first entangling the various definitions of artificial intelligence, scholars currently divide AI into strong AI and weak AI. Strong AI is universal, self-aware, autonomous, simply a person with the same intelligence, Star Wars R2-D2, Doraemon are strong artificial intelligence, have their own consciousness, can To make a decision for himself, strong AI is still an artificial intelligence in science fiction films. Weak AI goals are not that far-reaching, focusing on specific tasks that were previously only human or animal brains can do. The current breakthrough is still better for weaker artificial intelligence . Speech recognition, image classification, AlphaGo, etc. are all weak artificial intelligence systems to solve a specific task. Weak artificial intelligence needs a clear optimization goal: the optimization goal of speech recognition is the accuracy of character recognition, and the optimization goal of image classification is the accuracy of image recognition. Current AIs are almost all weak AIs. When there are clear quantifiable optimization goals, iterative learning approaches can be achieved through machine learning. In many of these areas, AI can approach or even exceed the human level. After talking about what kind of problems AI is suitable to solve, look at what AI is doing now . 1) The problem of unclear processing definition: The computer can operate according to precise instructions, and the execution is very fast, but the instructions are clear and explicit; 2) Perceiving: People can perceive the surrounding environment. People in the same conference room do not speak. A newly-entered person can feel whether they have experienced intense argument or happy discussions before, although the machine can be captured by different sensors. Specific figures of temperature, light, and humidity, these people can not be so accurate, but can not really combine these various information to reach the level of people. 3) Decision-making: Many decisions do not have precise rules. They involve the judge's intuition, mentality, and conjecture. These are very complicated for computers. These are necessary conditions for strong AI, so individuals think strong AI may not be realized in the foreseeable future. The current variety of so-called intelligent robots are more integrated with some weak AIs. They can integrate a system and can achieve specific tasks such as speech recognition, image recognition, speech synthesis, and chess, but these combinations can only solve each specific problem. The task does not constitute self-awareness and real thinking. Here are a few simple examples: When you are on the subway, you stand in the doorway. The person behind asks "when you get off the bus." The subtext is "If you do not, please give me a way." At this time, if you get off the bus, you need to answer "down"; you don't need to answer when you don't get off the bus. You just need to take the road off. When boys and girls talk about breaking up, girls will say more about “Who is she?†The above describes that when we do weak artificial intelligence, we need to define the optimization goal for each task. These optimization goals are all basic functions of human beings. Some functions, even animals, can be implemented and may perform better than humans. What is the goal of an intelligent person's optimization is more complex and may be a philosophical and sociological issue. Rights, money, the opposite sex, being recognized, the world together, and even the "emptiness," "Tao," "Nirvana," etc. in religion are the goals pursued by people. First of all, we cannot know how many goals we have in the end, and we can list the union of goals. It is not the sum total of all human pursuits. Secondly, only a part of these goals can be described by an optimization function, and then decomposed into various instructions. This can be considered as a weak AI system. For example, when preparing for an English test, the results of the test can be measured by scores, more words, more grammar, more fluent text, more correct pronunciation, and higher scores. In this way, we can use the goal of obtaining high scores and make a system of English exams. By allowing this system to take exams, we will achieve good results. In the next few years, the level of exams will exceed the best level of human beings. In returning to these goals, one person may pursue these goals at the same time; at different stages of life, or at specific times, these goals will also change: For most people when they are young, they will pursue power and money, and they will pursue their family and youth longer. When you are old, you pursue health and longevity. For each and every individual, education, surrounding people, and social changes will lead to dynamic changes in life goals. However, the impact of each event on specific optimization goals may not be known. Here is an interjective statement, although we can't be sure that every good book, a friend who is more cattle, and a more positive view of society, these will bring us much change, but continue to do it, people will certainly be better. It is impossible to quantify, exhaust, and integrate people's various pursuits. Not all optimization goals can be quantified. These are problems that the current AI framework cannot solve. From the conception of AI to the present, the development of AI is basically quantitative, and scientists do not jump into the overall framework. Therefore, individuals think that strong artificial intelligence will not be realized within several decades, and people do not need to worry about the emergence of strong AI and replace human beings. Many companies, large and small, and some experts put forward their own “AI is equivalent to the IQ of children†due to various considerations. In various weak AI areas, human beings can exceed the best level of human beings; in terms of perceptive decisions, AI is inferior to newborn infants, so now all the behaviors that claim the intelligence level of their AI products at the age of human beings are rogue. The above retired took a little rough idea and introduced some of the gods' dry goods. Professor Yang Qiang from Hong Kong University of Science and Technology gave a keynote speech on "Some Necessary Conditions for the Success of Artificial Intelligence." Teacher Yang Qiang believes that the algorithm model after deep learning has a three-layer structure: The first level is recursive deep learning (RNN); The second layer above this is a reinforcement learning learner (RL); The third layer is the migration of learning (TL), it can migrate an existing model to a new area. Deep learning and training requires a lot of data. Reinforcement learning also requires a lot of data for feedback. Using migration learning can effectively reduce the demand for data volume. For example, the experience of riding a bicycle can help you learn to ride a motorcycle. A person who can play badminton can learn tennis faster. This model has been applied to dialog systems and stock forecasting. Professor Yang Qiang believes that the current successful application of artificial intelligence is still in machine learning, finding patterns from data and replacing duplicated work. Finally, he gives five necessary conditions for his success of AI : Clear business model High quality big data Clear problem definition and field boundaries Understanding AI's cross-border talent Good at applications and algorithms Calculate ability. Professor Yong Yong of Microsoft Research Asia made a speech entitled "The Long Vision of Computer Vision from Perception to Cognition." In the speech, Professor Yong Yong mentioned that it is now possible to give a description based on the content of the input image, such as generating a "parking bicycle next to the river," and will continue to study the generation of descriptive text based on the video content. More exciting is the ability to answer natural language questions about image content. For example: On a muddy dirt road, what is dragging the carriage? The answer is that the horse is dragging the carriage. The method is to use a deep learning model of the text of a problem at the same time, a deep learning model of the image, and finally fuse the two models. When the author thinks that when listening to chorus once this year, different people sing different voices, and can quickly determine the sender of the voice based on his mouth movements and sounds. From the intuitive feeling, this should be the brain's input of continuous images and sounds together. , made a joint model identification. With the deepening of research, the joint input of future input sources such as voice, text, and images will surely make more interesting and practical products. In the same way, the last brave teacher also gave three key factors for the further development of computer vision: advances in machine learning algorithms, cooperation among scientists and practitioners in vertical fields, and high-quality mass data. (Professor Li Ming) Prof. Li Ming, a professor at the University of Waterloo and former teacher Lin Dekang from Google senior staff also introduced the keynote speeches on deep learning as natural language processing and automatic question and answer technology. At present, there are many startup companies that do AI, and various voice assistants and robot companies with various functions are surging. Each team has a strong academic background and technical ability. Good technology does not mean that you can make good products. There are still many other factors that need to be done to make a good product. Xiaomi co-founder Huang Jiangji said in his speech that Xiaomi’s view on artificial intelligence is: product + big data + machine learning . Manufacture various intelligent hardware products, generate high-quality big data, use machine learning to create artificial intelligence, and then apply it to various products to make products more intelligent. Sogou CTO Yang Hongtao believes that good products must meet three conditions: applause, selling, and making money . And he proposed that "only good products using AI technology can create value. One is to provide users with a good solution, and the second is to actually produce data so that they can iterate." This is consistent with Mr. Huang Jiangji’s view. "Desire before means." Using products and functions to satisfy users' desires in specific situations, instead of using existing technical means to find the user's needs, desires before the means, in turn, "taking a hammer to see where are all nails" is taking the wrong direction . Personally think that good artificial intelligence products must be able to land, have a carrier, can effectively improve people's lives. Take the author's out-of-home interview as an example, we have been committed to doing scene-based voice search, with independent speech recognition, semantic analysis, vertical search, and intelligent push technology. At the end of 14th, it released its own smart watch system Ticwear, and provided Android wear's voice search technology in China. Our own hardware product, Ticwatch, and smart vehicle-mounted products to be listed will ask the magic mirror. We hope that the artificial intelligence technology will land on the ground and facilitate the use of artificial intelligence technology to improve the quality of life. At the conference, an artificial intelligence & robot Top25 corporate trophy was also presented for the interview. It is possible that no one can predict the future of AI. As a practitioner, what we can do is down-to-earth AI technology and products, which will improve the production efficiency of society at large and free people from tedious and repeatable work. Hope AI can make life better. Audio Cable,Auxaudio Connector,3.5Mm Stereo Cable,Stereo Audio Cable CHANGZHOU LESEN ELECTRONICS TECHNOLOGY CO.,LTD , https://www.china-lesencable.com
Professor Yang Qiang of Hong Kong University of Science and Technology: Five Necessary Conditions for AI's Success
Executive Vice President of Microsoft Research Asia Yong Yong: Computer Vision From Perception to Cognition
