Seven tricky questions from teachers at the MTS School of Data Analysts

The first answer is the worst. On the contrary, you should not shoot sparrows with a cannon, but look for the most effective approach. For example, a regular filter can find a frame in an image.

The correct answer is the last one. The main application of Deep learning is in text (if regular expressions don’t help) and graph problems (by the way, we posted them in open source CoolGraph – library for quick start with GNN), in recommendations – both for DSSM architectures and transformer approaches, not to mention content embedding, for example, based on CLIP.

In all these cases, in order to get an acceptable result, you need to find data and train your model, and this is a complex and responsible process. Underestimate its importance it is forbidden.

Determining eye color from internet history is an inside joke, but there is a serious theory behind it. Yes, there are many different algorithms in basic ML. There are different classes of problems, and each of them has a dozen of its own algorithms. In our example, this is a multi-class classification, because the eye can be green, brown, gray and other color variations. That is, there are conditionally 4–5 answer options, but you only need to choose one. This problem is solved by different algorithms – and one of them is gradient boosting.

The correct answer is the first one. Docker does make deploying and managing applications easier with containers, but it doesn't automatically optimize code and doesn't run without an operating system. Packaging machine learning applications in Docker provides consistency in the environment, ensuring that models run the same from development to deployment.

With its help, you can include all the necessary dependencies in the container, such as specific versions of libraries and frameworks, and eliminate compatibility issues and dependencies on external libraries. This makes applications more portable, allowing them to be easily moved between different environments and platforms without re-configuration.

Students usually think that analysts only work with ML models, but at the school we teach analysts other technologies that they may encounter in their work. Knowing and understanding the principles of Docker, it will be easier to communicate with colleagues and you can create an MVP yourself, which greatly speeds up your work.

The correct answer here is the second one. MLOps does focus on automation and model lifecycle management, but is not limited to just hyperparameter tuning or eliminating human input entirely. As part of the MLOps unit at the School of Analysts, we dive into the world of integrating DevOps practices and machine learning processes and introduce key concepts of model lifecycle management.

Knowledge of tools and platforms such as Kubernetes for container orchestration, GitLab CI/CD for deployment automation, Airflow for regular model retraining, and version control systems will allow you to track changes to data and models.

Calibrating the probabilities of machine learning models plays a key role in ensuring the accuracy and reliability of predictions. In classification problems, calibration allows models to correctly produce probabilistic estimates. Without it, answers may be too confident or, conversely, give incorrect answers. Therefore, the correct answer is the first one.

There are several typical approaches to calibrating model probabilities. One of the most common methods is the Platt method, which uses logistic regression to convert model outputs into calibrated probabilities. Isotonic regression is also widely used, which is suitable for more complex cases and does not require assumptions about the functional form of the dependence. Both methods have their advantages and limitations, and the choice of approach depends on the specific problem and type of model.

Yes, the first two answers are indeed implementation options – and we will talk about them during the course. But in real life, even Near-Real-Time interaction is often redundant and expensive.

The most accurate answer is the last one, and during the course we will figure out why. The strength of the course is its broad scope. We will put into practice technologies for working with big data: MapReduce, Spark and many others. Let's see how optimal carry out analytical calculations using them.

Yes, in the field of data analysis and processing there is a role for a data engineer, and the question is more related to their competence. But in practice, such a role may simply not exist in the team, and experience with such tools will be a plus for the employee. There are also stories when you need to create a prototype, and the ability to make it yourself significantly speeds up and facilitates the work of both the analyst and his colleagues.