Statisticians are more concerned about the interpretability of the model, while machine learning experts are more concerned about the predictive ability of the model.
I don't think most statisticians agree. Statisticians are not more concerned with the interpretability of the model, but more concerned with the rigor of constructing and explaining the statistical model.
For example, machine learning, I predict that the probability of rain tomorrow is 0.8, and statisticians will say that my 95% probability of rain tomorrow falls between 0.7 and 0.9. We give confidence intervals, not just an interval, but tell people who see the conclusion how credible my prediction is. How to give this credibility? Using probability (probability theory). Probability theory is the basis of statistics, but it is closer to mathematics and more rigorous than statistics. Why does mathematics at the top of the food chain despise statistics? Because there are many inaccurate things in statistics, statistics also involve a lot of intuition.
Prediction is also the purpose of building statistical models most of the time. With the help of the existing data and facts, input them into the statistical model, add the corresponding random description, and output the prediction of the future. The world is not a function, the world is a function+a lot of randomness when realizing this function.
Peng Qi used this stem in his predecessor's answer, and statisticians denied that the machine learning model was actually wrong. He believes that statisticians look at the conclusion of the machine learning model like this: "If all the inputs of your Markov model are historical stock prices, then if the government promulgates a new bill prohibiting the development of X industry the next day, your judgment based on the favorable history of X industry is blind;
I don't think statisticians will deny your model like this. Our statisticians will regard the factor that "the government issued a new bill the next day to prohibit the development of X industry" as randomness. The value predicted by this Markov model is 90% credible, because there is a 65,438+00% possibility that the government will issue a new bill to prohibit the development of X industry.
There is a short story. When I was listening to the seminar of a statistician J.S. Marron (Professor UNC-Chapel Hill) in the department (IE Department), several of their professors joked: "People who think machine learning are making impossible predictions!" Everyone burst into laughter at that time. Scholars despise each other since ancient times. It is estimated that the chain of contempt in academic circles is interlocking and has existed since ancient times.
I don't know much about machine learning. I prefer to see the advantages of machine learning in this problem, not because people who engage in statistical theory are relatively low-key (I didn't find it on Zhihu. . . ), using the bragging skills of the industry to suppress classic statistics. Daniel, where are you all?
I think of another stalk. Isn't the book "Big Data Age" quite popular recently? I rolled my eyes and said, "In order to test these search words, Google processed 450 million different mathematical models. After comparing the predicted results with the actual influenza cases recorded by the US Centers for Disease Control in 2007 and 2008, the correlation between their predictions and official data is as high as 97%. " I was scared to cry by 450 million. . . I guess it's 450 million steps of stepwise regression screening to explain the variables? The correlation is as high as 97%. I guess more than n explanatory variables are used, resulting in R =97%? People who mix in the industry can really "blow". . .
I just read a link in the comment on the question. There is a sentence in it. Let me attach:
What is the difference between artificial intelligence, machine learning and statistical data mining?
Brendan O 'Connor's blog post statistics vs machine learning, spell it! The first draft was written in 2008, which may be related to the author's machine learning background. In the first draft, he mainly belittled statistics, and his thoughts were similar to [1]. He believes that machine learning is more about algorithm modeling than statistics, such as the maximum margin of SVM, decision tree and so on. In addition, he thinks machine learning is more practical. However, in 2009 10, he turned to abandon his original view that statistics is the real transaction: statistics, not machine learning, is the real transaction, but unfortunately it comes from supporters of bad marketing.
Some people commented on this sentence: "Statisticians are more concerned about the interpretability of the model, while machine learning experts are more concerned about the predictive ability of the model." In doubt.
When I first read this sentence, I couldn't agree with it. I glanced at the contents of answering the main speaker talk. According to my understanding of machine learning now (of course, I don't know much), I think this sentence should be like this:
Statisticians predict by establishing a strict statistical model, so it can give the credibility and confidence interval of the prediction. But because the complexity of the world has caused many problems, we can't establish a rigorous statistical model, and we don't know what the model or function behind a complex thing is, so the statistical model can't solve the problem of face recognition.
Part of machine learning only cares about the predictive ability of the model. In other words, I don't care what the principle is in my black box. I used a lot of parameters and adjusted them by manual violence until I got a model that can distinguish the faces of men and women well. Then I can use this model ~ ~ but theoretically I can't tell you how reliable this model is.
Why can machine learning do this? My personal guess is that in the process of drastic parameter adjustment, the intuitive and cognitive reasoning functions of the human brain are added, and the function of the human brain is much more powerful than that of the computer. When the learning model is established, the process of human learning is added. I don't know much about machine learning. This paragraph is just a brain hole. . )
Of course, there are also some parts that can build functions in complex problems. At this point, you can build a part that can build a function, open the conditions of the known part, and violently adjust the parameters of the remaining unknown parts. (I guess)
Supplement 2: I don't know much about the violent parameter adjustment of machine learning, but I think what @ Yay Yay said in the comments is to objectively summarize the process of parameter adjustment of machine learning, so that I can learn and carry it conveniently [I am a porter of Daniel's thoughts ~ ~ ~]
"Grid search, grid search is automatic violent parameter tuning. Set the running parameters to make the program run automatically. Run all night and get up the next day to see which combination of parameters has the best effect.
Perhaps the impression of violent tuning of machine learning is mostly due to deep learning, which is very popular now, but most people can't understand the principles, such as ReLU, Dropout and BatchNormalization. It is estimated that few people know why these are useful. They look like nine Niu Yi hairs, but they are useful anyway. If you want to use it well, you can only adjust the parameters by feeling.
But other more mature models, such as svm and ensemble, have theories about their high accuracy. Statistical theory supports why they have higher accuracy than linear classifiers, optimization theory enables them to obtain approximate solutions, and computer theory enables them to be applied in practice. "