Learn statistics, algorithms, how to apply the machine learning?

How statistics

kuankedashujushiyanshi· 2016-07-20 09:40:19

learning a time machine learning algorithm, and then look back machine learning problems, to learn to solve a good problem using machine in fact, is not an easy thing, especially to the entire model to explain , if you want to be able to explain the model well, it will be more difficult. Because the use of machine learning to deal with a practical problem is not only we have to learn how to use the machine learning algorithm, it is more important how to model the whole problem. I just started to learn intelligent computing, of course, an optimization problem how to modeling, modeling is completed after the solution is relatively simple. But in machine learning, the problem becomes more complicated, many machine learning books are about machine learning algorithms, like I had "a simple machine learning algorithm", focus on the algorithm, but the machine learning problems is not only a machine learning algorithm, and some other knowledge need we pay attention to.


, a machine learning problem.


Learning= Representation+Evalution+Optimization ( model , ( ) and optimization of algorithm "EN-US" style= "max-width: 100%; box-sizing: Border-box! Important; break-word word-wrap:! Important; ">) .



as shown above, machine learning is mainly composed of the above three parts. Corresponding to each part, there are the basic methods of every part, here some methods in my previous blog can be found in the introduction, and there will be in the future to add. The following is specific to each part to talk about.


1, said (or called: model:Representation

    that is mainly the modeling, referred to as a model. The main work is to complete the conversion: the actual problem into a computer can understand the problem, that is, we usually say that the modeling. Similar to the traditional algorithms in computer science, data structure, how to convert the actual problem into a computer can be expressed in the way. This part can see "simple and easy to learn the machine learning algorithm".


  • probability model: mainly is the conditional probability (

  • non probabilistic model: the main decision function (


2, evaluation (or called::Evalution


3, optimization (or called:Optimization


4, span" style= "max-width: 100%; font-size: 18px; box-sizing: important; word-wrap: border-box! Break-word! Important

summary max-width: 100%; min-height: 1em; white-space: normal; color: RGB (62, 62, 62); font-family: Arial; font-size: 14px; line-height: 26px; box-sizing: important; word-wrap: border-box! Break-word! Important; background-color: RGB (255, 255, 255);" >



two, what is the generalization ability of

Generalization ability), this is the success of machine learning a very important evaluation index. At the same time, the prediction error is the index to evaluate the generalization ability of the learning method.


1, cross-validation


  • simple cross validation: and above, a simple division of < /span>

  • S fold cross validation: data were randomly divided into S parts, including S-1 as training, the remaining 1 copies as verification, repeat, final selection model


2,

    "the final task of machine learning is to predict the actual use of good learning model data, which is the generalization ability of machine learning. We hope to be able to directly face the optimization function of test data in the optimization process, but in the actual process, we can obtain such a function, then use the optimization function in the training process to replace the real function. In the process of optimization, the local optimal solution is sometimes better than the global optimal solution.


three, why only the data are not



1, No Free Lunch (NFL, no free lunch)

    no free lunch theorem was first proposed in optimization theory, by Wolpert and Macerday jointly proposed. The conclusion of the theorem is that the performance of the optimization algorithm is equivalent to the mutual compensation of all the possible functions. The implication is that no other algorithm can be better than the linear enumeration of the search space or pure random search algorithm. Style= max-width: "

The lastest articles of kuankedashujushiyanshi

FOF theory and practice training (Shanghai January 7-8, _2017)

From the watch to underwear, big data 7 most unique applications

Take a dip in the analysis of mathematics.

Mid quantify the actual combat training camp third (Beijing Railway Station)...

The world famous mathematical conjecture, you know a few?

A large wave of optical fiber communication principle dynamic map, the United...

A rare sight! The beauty of the chemical structure is beyond your imagination.

Data visualization: how to use color to support the point of view