A good research problem – what does it mean exactly? What is it characterised by?

TT: A good research problem is crucial to any researcher’s work. It allows them to choose one issue they want to dedicate their time and energy to from hundreds of thousands of exciting issues. This requires them to define a strategy for how they will do science. A researcher has to decide whether he wants to choose a topic that is relevant, broadly understood but also very much under siege, or one that is more niche and original, although requiring a greater understanding of the research community? As a general rule, the work of formulating a research problem should start by asking a question, e.g. whether the performance of a neural network depends on the number of layers, or by posing a hypothesis, e.g. deep neural networks get better results than similarly constructed networks with less depth and then verifying it. A characteristic feature of a well-formulated research problem is that the answer to the question posed is not apparent and requires experimentation.

What are young researchers’ most common mistakes when looking for a research idea?

TT: I think that both young and old scientists, as a rule, make many mistakes. At the end of the day, that is what science is all about, to be wrong and then arrive at the truth based on those mistakes. When looking for an idea for a study, it is worth paying attention to several aspects: whether my idea is original, whether someone has not already come up with a similar method, and whether the problem actually posed is well motivated, e.g., real applications or limitations of existing methods, finally: is this topic fascinating to me? Often, as researchers, we try to follow the crowd and focus on popular issues or of interest to colleagues. It is worth following your own interests and taking advantage of your strengths – every one of us has a child’s curiosity, and it is worth getting carried away.

Where to look for inspiration?

TT: Life brings the best inspiration, especially the practical application of the methods described in the scientific articles. It often turns out that methods that are extruded on synthetic data sets and tested in laboratory conditions, in practice, are worthless. It is also helpful to periodically look at and participate in the most important events related to our research, especially scientific conferences on machine vision and machine learning methods. These are conferences such as CVPR, NeurIPS or ICML, where my co-authored papers have been published.