I happily announce that my new book titled “大規模計算時代の統計推論” has just come out. It’s available at Amazon.co.jp.
The book is a Japanese translation of the celebrated book “Computer Age Statistical Inference” by maestro statisticians Prof. B. Efron and Prof. Hastie. I supervised the translation project and translated a few chapters myself.
Although the topics covered are mostly basic, the description of those topics is deep and profound. What I was particularly interested in is the way statisticians look at machine learning. The authors say:
In the absence of optimality criteria, either frequentist or Bayesian, the prediction community grades algorithmic excellence on performance within a catalog of often-visited examples such as the spam and digits data sets of Chapters 17 and 18. (Epilogue, p.447)
They look as if they did not care about all kinds of development of empirical risk minimization and Bayesian learning (in the machine learning sense). Statistical machine learning does have criteria for optimality. If not, why would machine learning researchers prefer to write down their algorithms in terms of an optimization problem?
The authors’ idea of “optimality” looks very strict, which is a good thing per se. Statisticians look like living in a country where strictly separated tribes are fighting each other. Something coming from the outside of the country is not even worth considering. This probably wrong impression seems to explain what I have experienced many times when talking to statisticians. The good news is, however, that more and more young statisticians appear to be freer from the old-school statistician’s way of thinking.