How Machine Learning Works and How to Receive an Award by Yandex

15:52, February 20, 2019

The Head of Yandex Machine Intelligence Lab gave a lecture at UrFU

Today machines learn how to compile weather forecast, react to illogical user search queries, track current road traffic conditions, pick relevant music and tickets, talk to people, analyze particle behavior in Large Hadron Collider and many more.

Alexander Krainov, the Head of Yandex Machine Intelligence Lab spoke about how search queries are formed, how the “decision trees” method works, and how machines pass tests. The expert gave a lecture, during which, among other things, he told UrFU students about how the company trains and retrains the search engine to show relevant links for a huge variety of requests.

‘Machine learning is when we choose a formula and find the appropriate parameters for this formula, going through them. The formula can be very simple: let's say there are only two factors: the presence of a certain word on the page and the number of links to the page. And we consider the sum of these factors and that's it,’ Alexander Krainov explains.

However, the formulas used in search engine training are rather complex.

‘We compare “warm and soft” in a search query: the quality of clicks, the presence of words, the user’s region, the update frequency, and so on,’ the developer says. ‘And here the method of "decision trees" comes to play. We have many "weak" trees, and each of them looks like this: we take some factor and divide it into something (for example, by the number of links: whether there were more than ten thousand or less). After that, we take another criterion, for example, whether the language is English or Russian. The tree is “branching out”, and, lastly, some decision is shown.’

In other words, the machine selects nodes in a variety of such “trees” in the process of learning automatically. The result is a list that summarizes the outputs from all the "trees".

Other Yandex services are trained in a similar way: weather, music, translator, self-driving car, Alice (virtual assistant).

Alexander also told the students that on January 30 Yandex has launched the Ilya Segalovich Award. It is a scientific award named after one of the company's founders. The scholarship of 350 thousand rubles (excluding personal income tax) will be awarded to undergraduate and graduate students for achievements in the following fields:

• Machine Learning
• Computer Vision
• Information Search and Data Analysis
• Natural Language Processing and Machine Translation
• Speech Synthesis and Speech Recognition

The recipient of the award will be paid 350,000 rubles (excluding personal income tax) over the course of ten months starting from September 2019 until June 2020 in monthly installments of 35,000 rubles each.

The laureates will also be granted a trip to an international conference on artificial intelligence, a personal mentor and an invitation to a scientific internship in the research department of Yandex. The winners’ academic advisors will receive a cash premium of 700 thousand rubles. The award ceremony will be held in April.

Applications are accepted until February 28.