Mikhail Okunev — about trends in machine learning

HSE-Perm has hosted the first seminar of the GAMES research group in this academic year. The key speaker Mikhail Okunev told the audience about machine learning through the example of his work in Facebook, Google and Microsoft.

The seminar format was open dialogue. Mikhail explained how he became interested in this topic, in which projects connected with machine learning he took part. “I would like to speak about different tasks, about what is going on in the machine industry now, to share my experience, and also to speak about the job of a machine learning engineer and how to become such specialist,” the expert noted.

For example, Mikhail was working for a long time in Microsoft division that was developing the Bing search system, in the Facebook team etc. As part of his job the expert was solving different tasks with the methods of machine learning: he was detecting spam and fraud, analyzing tonality of texts, searching unreadable duplicates in a huge data base and performing other tasks. He shared his experience, having analysed different approaches in detail. For example, Mikhail was developing an algorithm for ranging and highlighting of popular comments in the Facebook feed. He had to take into account many criteria which included the quality of a comment, number of likes, absence of spam etc. “Such tasks fall within standard machine learning,” Mikhail underlined, “The main thing here is the accuracy of prediction”. The expert told about the methods of gradient boosting which help to enhance initially imperfect mathematical model and increase the accuracy of prediction up to 90%.

Mikhail also compared econometric models and methods of numerical optimization. For example, econometric models take into account thousands of conditions. As a rule, they are simple, describe the data well, focus on interrelation and casual effects. Methods of numerical optimization are oriented for creation of complex non-linear models based on training set of data; the accuracy of such models is checked on a test set and then run on new data. The participants of the seminar in general agreed with such comparison of methods.

In the end of the meeting Mikhail noted that the main purpose in machine learning today is to make its usage common for non-specialists. Today machine learning is intelligible for a professional who is good in mathematics and IT.