Machine learning is used by companies that want to offer customers a useful product or increase the efficiency of their business. This is an opinion of a researcher in the artificial intelligence laboratory at Sberbank Dmitry Babayev. The same phrase explains why companies are interested in AI and Data Science technologies. Learn more in our interview.
Dmitry Babayev was in charge of the Data Science Department at MTS, as well as developed an algorithm for auto-completion of queries in Yandex search engine. Currently, he is engaged in research in the field of deep training for the banking sector.
Interviewer: AI Conference (AI)
Respondent: Dmitry Babayev (D.B.)
AI: Dmitry, what is the process of neural network training?
D.B.: With the help of gradient descent. It is used for many machine learning models, not just neural networks. Even for popular methods of linear and logistic regression. Neural networks apply one of its versions – stochastic gradient descent, which allows not to load into memory all the data that is necessary for learning. It is possible to use many servers in a distributed version of the algorithm, which can significantly speed up the resource-intensive learning process of the neural network.
AI: What are the challenges faced during machine learning?
D.B.: There are a lot of them. There are problems that depend on the choice of method, but there are also common challenges to all machine learning. For example, there is a problem of retraining, which occurs when using any algorithm of machine learning. Simpler it can be described as follows: the algorithm remembers the training sample instead of building generalizations on it. This is a typical problem for the whole machine learning.
In particular, neural networks have two of them:
The 1st one is the amount of data necessary for their training. A neural network is a model that has so many parameters, so it needs to show a huge number of examples for learning. In most cases, each of them must be prepared by a person. Thus, it is a problem of labor intensity of creating a sample for training.
The 2nd challenge is the number of hyperparameters. Each neural network has its own architecture, layers, types and order. Each time to find the right one is a special art, and each time it requires the intuition of the researcher, as well as the trial and error method.
AI: The technology of machine learning can be found, first of all, in banking sector and telecommunications. In what other areas can it be applied?
D.B.: Practically in all, where there is data from which you can benefit. Every company want to offer customers something useful and don’t want to offer useless things. And the technology of machine learning can help in it.
Internet companies have long and actively used the technology of machine learning, for example, in systems of personal content recommendation.
In traditional industries, there are success stories of the machine learning technology application. For example, together with Yandex Data Factory, machine learning was introduced in the steel industry for the correct selection of the alloy compositions, achieving decent savings.
AI: Which Russian companies has their own AI and Big Data laboratories?
D.B.: Only a few Russian companies has such labs, since usually it is connected with academic issues. Yandex does have the research department.
There are international companies that open their research centers in Russia. Samsung, for example. I don’t know how much this will suit Russian examples ... Perhaps Alibaba will open such a center in Russia.
AI: Is it because they need our specialists?
D.B.: Yes, indeed. Russia has a very strong mathematical education. The country has a lot of high-quality specialists in Data Science and AI that is why companies are opening research centers in the Russian Federation.
AI: How do you think: what percentage of operations will be performed by artificial intelligence in a year or five?
D.B.: It is difficult to look into the future, because the ubiquitous use of AI is influenced by two aspects: technological (what innovation is capable of) and social (what society is ready for). A good example is the technology of unmanned vehicles: there is an opinion that in the US today it is possible to replace all long-haul drivers on the routes between cities with machines. In total, it is about seven million workspaces, the disappearance of which will lead to social consequences which can be hardly predicted in a country with a population of about 300 million people.
AI: Who should think about mastering a new specialty in addition to long-haul drivers? In other words, which jobs will be replaced by the artificial intelligence first of all?
D.B.: Yet another difficult question. I think that today the machine learning technology cannot fully exclude people, it perfectly works together with people. For example, if earlier an advertising agency had about 100 managers of advertising campaigns, after the introduction of the algorithm of automatic selection of the target audience there will be 10 of them, but very good ones. They will process edge cases where AI fails.
It will first appear where there are many routine operations that can be described algorithmically; where a person works under a pattern, without creativity. If there is a social component in the work, the replacement of people with the AI won’t take place soon, despite the chatbots and dialogue systems. Definitely, not in the next 5 years, but it is only my opinion.
If we talk about robots (traditional ones), then, believe it or not, a robotic cleaner will not appear soon. It is a complex technology, the robot requires orientation in space, navigation and precise actions. That's what I dream of automating at home (smiling - author's note). Or a robotic cook. You come home from work, and it cooks and sets out the table itself. It would be nice.
AI: And what about robotic vacuum cleaners?
D.B.: It is quite a simple thing; I have one at home. It can only slide on the floor, and the amount of dust is reduced. And that’s it. But it cannot dust shelves for instance.
Do you agree with the forecasts of the expert? Come to AI Conference in order to talk to Dmitry and other speakers.