Some artificial intelligence as the next outlet, but I would say it is too early

Google (microblogging) Artificial Intelligence (AI) applications AlphaGo beat Shishi let the world marvel at some industry play Frozen Games insiders and therefore the depth learning technology touted as "the human brain simulation." But in fact, machine learning is far from being a magic bottle was released from the "monster." It's just a mathematical algorithm, understanding of human intelligence, to establish a comparable level AI and human step forward in the process. Deep learning that math Deep learning is quickly "swallowed" AI, but do not put this in the ascendant AI technology exaggerated. Well-known British writer Arthur Clarke (Arthur C. Clarke) once said: "Any sufficiently advanced technology is no different from magic." Deep learning course is an advanced technique that can identify objects in pictures and face, recognize spoken words to translate one language into another language, and even human beat top players in the game of Go. However, the depth of learning not to the extent of "magic" of. With Google, Facebook and Microsoft and other technology giants continue to integrate the technology into daily online services, and the world is still as Google AlphaGo beat Shishi occasion and surprised some industry insiders began to depth learning technology touted as "people brain simulation. " In fact, the machine learning is a simple mathematical calculation, but only large-scale computing. In fact, the depth calculation is an algorithm based on neural network data to adjust. So in the end is what these words mean? Let us explain: A neural network is a computer program (and inspired by the structure of the brain), which comprises a large number of nodes connected to each other (or "neurons"), numeric input each node received a simple function calculation (such as sum). Here the "node" is simpler than neurons in the brain, and their number is much lower than the number of neurons in the brain. Depth study only reinforces these connections in the neural network nodes. Deep learning machine learning is a subdomain, machine learning is a very active field of AI research branch. In theory, machine learning is the approximation function (approximating functions) method is based on a collection of data points. For example, if the arrangement is a set of numbers "2,4,6 ...", then the machine will be able to predict the fourth digit should be "8", the fifth number should be "10." The formula is 2X, X representative of the sort of position. This algorithm is very broad application space, for example, self-driving cars, voice recognition and prediction ticket price fluctuations and other aspects are outstanding. In a sense, the depth of learning is not unique, no rules can be found. Any machine learning systems, whether they are "deep" learning, consists of the following basic elements: 1. Run elements: part of the system to take action. For example, the part responsible for playing chess in the chess game. 2. The objective function: The learning function. For example, the position of the board game Go or go to map sub-selection. 3. Training: Data points marked a collection, for the approximation of the objective function. For example, the board position in chess game collection, where each location is marked by the human expert in that position to go sub-selection. 4. The data show: each data point will usually be represented as a vector of predetermined variables. For example, the chess set on each piece of position. 5. Learning Algorithm: Calculate the objective function approximation algorithm based on training data. 6. hypothesis space: spatial learning algorithm may be considered a function. This structure can adapt to all take machine learning methods, ranging from simple linear regression to complex deep learning algorithms. Technically, we are referring to is the supervised learning (supervised learning), where each data point has made human mark. The data has not been marked, that is unsupervised learning, it is necessary to deal with many difficulties. If parts of the data is marked, then the semi-supervised learning. It should be noted that five of the front part of the machine learning framework are manually entered, realize that this is very important. Construction of a human programmer where each element, but does not control the machine learning program. In fact, programmers usually analyze the behavior of this learning process, after it was found not perfect, will play Barbie Cooking Games manually modify one or more elements. This is a very hard work, before reaching the desired level, it may take years, or even longer duplication of work. Help humans We will find that the ability to learn a program will be severely restricted this architecture. Rather: 1. The learning process can not modify any part of the architecture. 2. The self-learning program can not be modified. 3. The learning process can not "learn" function assumes that the space outside. Because of this, such as AlphaGo learning process, without human help, and can not learn chess or checkers. In addition, if not after a lot of specialized training, most programmers are unable to successfully modify the machine learning system. Even well-trained data scientists, but also it requires a lot of time and resources to successfully establish a successful machine learning system. Design and Implementation AlphaGo system requires more than 30 million drawn from the training sample on the Internet, as well as a large team of researchers and engineers of years of hard work. In fact, only the lifting AlphaGo beat European champion Go Fan Hui levels to defeat Li Shishi level, but also take several months to work. In addition, AlphaGo also uses a series of machine learning method is known as "enhanced learning" through continuous selection operation, and observe the results in order to obtain the greatest chance of winning. Strengthen the learning process, the training data is not a "pre-labeled" input. Instead, the study program is to provide a "reward function", that is a different state assigned different rewards. Reinforcement learning through the implementation of certain actions, and observe incentives to obtain training data, machine learning analysis in this article also applies to intensive play Angela Games study, which is still limited in its objective function, performance data and assumptions space. Possibility space Very unusual learning power brought significant results, we usually call it "evolution." But it needs to point out that we must be aware of the difference between the evolutionary process of natural selection, and computer program to simulate the processes. A computer program to simulate a process called genetic algorithm (Genetic algorithms), the current method is not particularly successful. GA modified form of "life style", and this form of expression is very large. For example, the human genome contains an estimated 10 billion bits of information, which means that the number of possible human DNA sequence is a power of 2, 1 billion. Exploration costs on this scale space required is very high, but the topology of the space it simply can not be found, a suitable algorithm. By contrast, the possibility of much smaller space of Go, explore the use of machine learning methods will be much easier. In order to successfully define an objective function, will be living in a task becomes a simple optimization problem, computer scientists, researchers and statisticians It took ten years. However, many performance problems before the machine is operable form, but also the need for more analysis. For example, how to write the language the machine can understand the meaning of a word? As MIT professor Gerald Sussman (Gerald Sussman) puts it: "If the expression does not come out, will not learn." In this case, select the appropriate method of expression can not do, let alone solve the problem. Therefore, the depth of learning (or broad sense is "machine learning") has proven to be a powerful method of AI, but the current machine learning approach also requires a lot of human intervention to some of the issues is presented as a machine operable form. After that, also requires a lot of skill and time to define these issues repeatedly, until these issues were eventually resolved machine. The most important is that the process is further defined in a very narrow range, the machine is very low autonomous authority. And people do not like, AI does not have autonomy. Thus, machine learning is far from being a magic bottle was released from the "monster." Instead, it's just smart to understand, to establish a solid step with play Dora Games human-level AI equivalent process in the step.