Coding, Math, and Business Understanding: How We Train Future AI Researchers and Entrepreneurs
The main problem with AI training is the lack of systematic training
Let's start with a short excursion into history. Narrowly specialized sciences in the field of teaching artificial intelligence have not existed since the 2010s. But the topic of artificial intelligence is much older than it seems to us. It originated in 1954. And until the second winter of artificial intelligence – the beginning of the 1990s – it existed as a separate science of artificial intelligence. Not that it was taught in Russia, but there were certain institutes abroad that really did study this science. They included semantic, connectionist methods – the latter just grew into neural network methods. Back then, mathematical statistics did not get anywhere. Now it is the main base and apparatus for building models, but back then they were looking for other ways, expert systems, and so on.
During the second AI winter, when PCs appeared and investments in AI decreased, knowledge about artificial intelligence moved into computer science. By the 2010s, all this had transformed into Data Science. At that time, the approach to artificial intelligence switched completely to connectionist mathematical statistics. That is, large amounts of data were used to build mathematical models of varying complexity and depth – and these models made it possible to solve problems related to artificial intelligence. This is where we are now.
Fundamental education in this area is still just emerging – we are trying to help it in this. And 20 years ago, there was no systematic training of this kind at all.
It is important to understand that working with artificial intelligence predetermines the unity of three entities:
Coding. Programmable Data Science is rated significantly worse than developmental Data Science, because it is still a tool.
Mathematical apparatus. Mathematical statistics, probability theory and formulas-formulas-formulas.
Business understanding. This is precisely the weakest function for programmers. But it is important to pump it up in order to understand what to apply data to, navigate semantic stories, etc.
By and large, the first data scientists hardly left business. Usually they took two paths. The first was to retrain from development. Then it was necessary to gain expertise in the field of mathematical statistics and other mathematics. The second was to retrain from mathematicians. All universities that fundamentally prepare mathematicians well are the basis for the emergence of data scientists to this day.
In general, there were two sources – programming or fundamental mathematical education. And then all this knowledge needed to be supplemented: take online courses, search for information, check these sources, exchange experiences in communities – for example, in the same Open Data Science. This is where problems could arise, because the information was not systematic and it took a lot of time to find it. That's how it was for me.
My main role is to translate from technical to human and back. That is, I need to be able to interpret. There are few such people in the industry: most often, if you are a techie, then you are a techie. And during my studies, my problems were mainly related to this. My basic education is Baumanka, no one taught machine learning there. And I needed both developers and data scientists to understand me. I did not always have enough technical knowledge for this. So I had to do a lot of manual work: search for materials, extract information from more experienced comrades, monitor communities, see what is suitable, what is not, try and try again. I looked for free or paid international courses – some on Coursera, some on DataCamp, some somewhere else. All this took a huge amount of time. We suffered a lot with this back then.
When I entered the profession of a machine learning engineer, I only had a mathematical background. I started programming back in the distant past at university in BASIC and a little in Pascal, then happily abandoned it. And when I was already interested in data science, it was not easy for me to find a good course in Python. Even more difficult – in SQL, NumPy, pandas. In general, it was difficult to immediately understand why all this is needed in machine learning. I collected all the information bit by bit from remote corners of the Internet, from acquaintances and friends, looked for good communities. I really lacked like-minded people who would be on the same level as me.
Difficulties arose with understanding machine learning models and interpreting them. Machine models are different, and there are a lot of them: linear models, trees, boosting, neural networks. How to navigate, where to apply what?
Then the interviews started, where I was given algorithmic tasks. At that time, I didn’t understand at all why they were asking me about this, where algorithms are used in machine learning. I realized all this later – when I had to interview the guys myself.
In the Master's program, we provide an understanding of how and where AI works
To explain where AI researchers are needed today, let's get a little academic. Let's start with a definition: what is research anyway? It can be divided into two types:
They have fundamental differences.
The purpose of industrial research — get a result. That is, we take something ready and try to apply it, get a better result. If you need it to whistle better, we hang a whistle.
The purpose of scientific research — to gain new knowledge. And this is a completely different story. Here you gain an understanding of what and why works this way. You gain some understanding of the processes.
Often, research is understood as industrial research. But in the Master's program, we primarily prepare those who will engage in scientific research, that is, obtain new knowledge. Such people are now needed in three types of institutions:
Scientific institutes: Institute of System Programming (ISP) of the Russian Academy of Sciences, Computing Center of the Russian Academy of Sciences, etc. For example, at ISP, they are currently engaged in research in our field and also in applied developments.
Universities: Phystech, HSE, Skoltech, ITMO, MSU, NSU, and so on. These are primarily educational institutions, so there are many more students – as a result, employees have a greater workload in the form of teaching. That is, if a person goes to work at a university, he will be engaged not only in science, but also in teaching. Not everyone likes this, and you need to be prepared for this.
Industry. Some companies conduct scientific research to improve their products – and AI researchers are needed here too.
What are the risks of self-taught people?
The most common trap that haunts self-taught people is the incorrect understanding of the context and use of terms. We even had a dispute the other day: is it cosine proximity or cosine distance? In practice, they say either way – but they essentially understand what is happening. My example is local, but such discrepancies can really lead to problems.
If you misremember the terms during self-study, you may be considered uneducated, unable to analyze what is happening. You will not even pass the interview. If you do not speak my language, I am not ready to work with you. Within the team, this will inevitably lead to problems like “I was given a task, I understood it incorrectly.”
A real-life example: in the fall, I talked to several artificial intelligence startups from San Francisco. We were discussing a platform to help sales managers. In Russian, “sales manager” sounds normal – it's a manager, a salesperson. But in English, it means something different – they mean the head of the sales department. As a result, they stubbornly did not understand me. The same thing happens to an uneducated data scientist in business, who simply cannot correctly interpret what they are told. That is, they cannot use the language that is accepted in the company and in the industry as a whole.
What other traps can self-taught people fall into:
To draw wrong conclusions from wrong facts.
Lose one of the three entities – remember, I talked about them above? Most often, it is business understanding. That is, you may have knowledge, but at the same time you may not be able to use it in business.
Not to highlight the meanings. You can conditionally take basic courses on the Internet and start training models. But a data scientist needs to understand mathematics, business and the meanings that are embedded in what you call in the code. If you only engage in self-education, there is a very high risk of not getting this knowledge and meanings.
Develop slowly. Of course, when you learn on your own, the learning curve suffers. When we search for knowledge on our own, the speed of development decreases. Searching and understanding what is suitable and what is not, testing and figuring it out – this takes a long time. Of course, self-taught people in any field can be stars, but how long it took is another story.
Master's degree accelerates students' entry into industry
To maintain your expertise in AI, you need to constantly study scientific articles and reports from international top conferences. This is the main source of knowledge, because everything changes very quickly, and no education can replace it. But at one time we managed to collect bumps with the search for information, and we got ideas on how to make this process easier for future specialists.
We came up with the idea of a Master's degree to make learning systematic and teach things that are difficult to master on your own. For example, everyone can read scientific articles, but not everyone does it correctly. We will teach our students to read scientific articles in such a way as to extract the correct meanings from them.
But the main thing is that a master's degree helps to combine the three entities I mentioned above in one education. Graduates will be strong in coding and in the mathematical apparatus, and at the same time understand business. At the dawn of the 2010s, everything was different: a person who was more or less close to you came to work for a company. They told him: “You will do this, and now this and this.” And over time, he honed his skills. But now the pace in companies is fast, and there is simply no time for this.
30 places in the Master's program are financed by MTS, meaning that education will be free for 30 people. At the same time, graduates can choose any career path, it is not necessary to go to work at MTS.
We often have unprepared guys come to our interviews. They take online courses and come to work for us, but they have no understanding of how to apply their knowledge in practice. For example, someone cannot solve a problem on Bayes' theorem and does not understand how probabilistic approaches work in ML in general. Such an employee simply will not be able to plan and conduct A/B tests, validate the results of their experiments, and so on. And we do not always have enough time to help such guys fully get comfortable.
Studying on your own is, of course, also good, and is always welcome in IT. But will you be able to ask a question to the author of a YouTube video? Or work closely with an entrepreneur you like for several months: attend their course, ask questions, ask for advice? Communicating with live teachers in real time is always much more effective.
Colleagues have also encountered this. And then an idea came to mind: why not open our own master's program that will simultaneously train both researchers and entrepreneurs in the field of AI? On the one hand, we will give students fundamental knowledge, and on the other, an understanding of how to use it in practice right now. And we will have cool people, hurray!
The Master's degree will give an understanding of why each formula is needed both in ML and in business. And the part of the program where we talk about startups will allow you to dive into the world of business and get closer to business customers. This is especially valuable if students decide to go to work in the industry.
We want to train guys who are focused specifically on AI research, who will go into science, open their own startups or work in commercial companies – in the RnD team of MTS, for example. And since we ourselves are interested in personnel, it is important for us that the guys quickly join the industry. That is why our form of study is evening. So it is quite possible to combine a master's degree with an internship and a full-time job.
How to teach entrepreneurship and work in science at the same time?
We will give students knowledge and immediately immerse them in practice. When a master's program is created in partnership with a company, as in our case, it is always more effective. Students will immediately begin to practice and create their own projects. Already during their studies, they will receive a basis for their own research and launching startups in the field of AI. And in the first semester of the second year of study, you can take advantage of the exchange program from HSE and gain experience abroad.
You can find out more about the program and admission trajectory here. You can apply submit until July 25.
It is important for us that during their Master's studies, students gain comprehensive experience and feel out what is closer to them. For this purpose, the program includes presentations by a wide range of specialists, including active entrepreneurs. Students will learn not only large language models, data analysis, generative neural networks, etc., but will also learn how to properly attract business angels to their projects, how to recruit a team for a startup, and how to choose partners wisely. As active experts in the field, we will share where we and our colleagues most often made mistakes.
We will also teach how to write scientific articles, which is essential in scientific work. Vadim Strizhov, one of the mentors, will conduct seminars on writing articles, and he has been in this field for over 30 years. I am sure that with his expertise he can give students something that no other master's program can give.
The superpower of our Master's program is our teachers. These are people who are currently researching artificial intelligence. They are all passionate about science, have been working with AI for many years, and can directly impart in-demand knowledge to students. For example, we have a teacher who has been in NLP for 11 years, and he will teach this course. Are there many people in Russia who can talk about NLP in a quality and engaging manner? I know three at most. In general, I would be happy to listen to our teachers myself.
It turns out that already during their studies, students will closely interact with the industry. They will be able to learn from the example of teachers and, if desired, go to work in the same field. Those who prove themselves well can receive an offer from MTS already in their second year.
Master's students can start working while still studying
The field of data analysis is developing very quickly now. Analysts are sought with knowledge of ML, so that they can roll out a simple model, quickly calculate something, predict. Students will already be able to do all this and will be able to get a job in a company – for example, as industrial experts. That's why our master's program is evening: classes start after six in the evening or on Saturday afternoon. So a student can work his 40 hours a week fully.
For the guys who show themselves well, we will organize internships at MTS. In essence, we will take them by the hand, show them how things work. And such integration into the field is a huge advantage of our master's program.
Master's graduates, for example, will be able to create classic churn models if they want to go to a bank. Or work with large language models. Or maybe they will realize that they like science, go to graduate school and do scientific research at academic institutes and research universities.
Graduates of the Master's program will have a diploma from HSE, a diploma from the Faculty of Computer Science. Even if there is no diploma yet, but a person is already studying at HSE in a Master's program from MTS – this is already a huge plus for the employer.
Now one of the teams at MTS AI is researching large language models for code. The guys are making a programmer's assistant, that is, a technology that will:
predict the next word when a person is writing code;
good at generating unit tests from a piece of code;
generate comments well.
These are the tasks we are working on right now, and graduates of the Master's program can very well be junior colleagues in the team. There will be even more such tasks ahead, because the field is developing. Go for it!
![[Проект вуза] Psychoanalytic coaching that works](https://prog.world/wp-content/uploads/2023/02/cc3f6051be2b047be8c49ffdc7d6f61e-768x403.png)
