before writing a "from Alphago talk about general-purpose AI Design" article, which discusses the general-purpose artificial Intelligence design framework. Then someone asked me how to do the special-shaped AI, because they have done image recognition and speech recognition two AI projects, a little bit of experience, so today based on these two AI projects, to talk about the general specific type of artificial intelligence design and design requirements. Again, the specific type of artificial intelligence needs a variety of specific implementation methods are different, here only according to the AI project I have done, talk about the design of specific types of artificial intelligence basic requirements and design ideas, do not guarantee the application of other artificial intelligence field, also does not involve coding. In fact, the structure and thinking clearly, how to encode, with what computer language, is just a process.
Come on ! To do a specific type of AI must have a strong enough environment for storing and computing data, and this environment, with the increase in storage and computation, can synchronize linear growth, can provide this scale of magnitude of the environment, is now only big data. Some people may say that using the database line is not possible, here can tell you clearly, the database is really not good. Take Oracle for example, in Oracle, storage hundreds of GB is the limit, and then the performance becomes very poor. However, these data are in the artificial intelligence environment, it may be just a calculation of the amount of data, the two are not at all a magnitude. Now we do artificial intelligence testing system, the amount of data stored is close to the PB level, such data volume, the database is not done at all.
go back to the computer nature to talk about. No matter how magical the artificial intelligence, in fact, the computer is just a machine, can never be a machine, it can not have people's thinking ability, all the implementation of artificial intelligence is through the computer powerful data calculation to obtain, so, after having strong enough data environment, There is also a need for enough data to enter this segment: Data acquisition. There are many ways to collect artificial intelligence, the simplest and most convenient source is the Internet. But the data acquisition is a quite long process, only after accumulating enough data to be able to do artificial intelligence calculation, otherwise it can not achieve the accuracy and accuracy requirements. Strictly speaking, this job is not a technical work, is a physical activity, but also to add time consumption. Take our two artificial intelligence project, accumulate to today this data volume, before and after spent more than two years time, the data source has the Internet, also has the partner to provide. So, in terms of data volume, it is not a simple matter to want to do artificial intelligence.
achieve the above two goals, the rest is the most critical part of AI: deep learning. I see articles on the Internet, and some people call this piece "self-learning", or "smart statistics". But no matter how it is understood, its essence is based on data analysis and data calculation. In order to facilitate accurate artificial intelligence calculation, in the process of data acquisition, the input data, before entering the data storage environment, will be divided into smaller data tuples according to different specific requirements. These small groups of data are often labeled with different labels to indicate the various attributes that these data tuples have.
There are a lot of data errors that we call "data noise" during the creation of the data tuple. To eliminate the data noise work, it is definitely a person to do, because the computer is not aware of such a right and wrong. Operators classify them in the "wrong" piece. When the "wrong" data accumulates to a certain amount of time, because there is already a recognizable basis, the work can be done instead of people. In this way, after continuous purification, the amount of data will be more and more, precision will become more and more high, which will be the basis of artificial intelligence "deep learning".
the second step in deep learning is to create an association for the data tuple. This association is meshed and increases in geometry and extends as the amount of data grows. At present, the calculation of artificial intelligence is so large, that is, because of the association between the data tuple too much, resulting in a variety of possible factors too much, resulting in the computer in the calculation process, to identify and filter one by one, to achieve quweicunzhen effect. and the elimination of data noise, to set up the work of the data tuple, initially can only rely on artificial to achieve, when accumulated to a certain extent, the computer will replace people's work. In the outside world, this time the computer seems to understand the behavior of people, with a certain degree of intelligence.
take an example to illustrate the process of artificial intelligence: face recognition. This is a branch of image recognition, but also the most technically mature AI project. Its principle and process is this: no matter what the race, the positive facial contour is basically the same. Based on this premise, before face recognition, the image software will scan a lot of face contour (vector processing), while the face of the eye, mouth, nose and other organs are also scanned, including the position of each organ in the face of such parameters, and then labeled, annotated properties, and they are stored in the form of numbers. There are other personality characteristics, such as skin color, lips, eye color, high nose or collapsed nose, etc., also labeled labeled Attributes, as a more detailed identification options. When we say that XX and xx resemble each other, the computer does this: scan their facial contours, compare them to the face parameters that the computer already exists, and, if their facial contours are consistent with most of the parameters, basically make sure that this is the image of the two faces. Compare the data of two faces and set a similar value for different organs (80%,90% or ... If the overlapping parts are within the scope of the requirement, then the computer is judged to be "elephant", otherwise it is not like.
such facial images, in the case of insufficient basic data, can easily occur recognition errors
Finally, we summarize the specific type of AI with two articles:
1. Big Data + big calculation is the basic condition of realizing artificial intelligence.
2. The role of data correlation in AI is important, and all AI relies on it to achieve it.
In addition, when it comes to coding rules, in fact, every kind of AI will be applied to different disciplines in different fields of technology, it is not easy to achieve, here is also very humble writing, mainly do not want to give you the difficulty of reading and understanding, you and the GU read it.
On the design of special-shaped artificial intelligence