Gdańsk Tech Scientist with Publications at the European Conference on Computer Vision in Milan | Gdańsk University of Technology

Sebastian Cygert, PhD, Eng., who also works as a postdoc at IDEAS NCBR, conducts research on algorithms taught in continual learning mode, with an emphasis on their safe operation in changing conditions. His work concerns the creation of AI systems that can independently adapt to the environment without the need for constant human intervention.

Sebastian Cygert is the main co-investigator of the project “MLAdapt - Self-adaptive machine learning”, financed by the National Science Center under the Sonata program. The aim of the project is to develop self-adaptive algorithms that allow AI models to adapt to new conditions autonomously.

– Currently, the dominant AI paradigm is based on static models, such as foundation models, whose parameters remain constant over time. However, in some applications, especially dynamic ones, the ability to adapt independently may prove crucial. However, this innovative approach is associated with numerous research challenges that still require in-depth basic research – explains the scientist.

Continual learning

The potential applications of such adaptive algorithms are wide, from financial models that respond to market changes, to algorithms for analyzing medical images that must adapt to new diseases, such as COVID-19, and changes in existing ones. Similar solutions are also used in the analysis of social media, which are constantly evolving under the influence of new trends.

In the article “Category Adaptation Meets Projected Distillation in Generalized Continual Category Discovery”, the authors examine the problem of discovering new classes of objects in a continuous manner (Continual Category Discovery). They propose an innovative method that increases the accuracy of models in such tasks. The team's contribution consists in combining two existing methods and presenting their effectiveness in a new context.

In other works, a new method of effective knowledge accumulation for continual learning algorithms was presented and analytical studies were conducted on the effectiveness of individual components of the architecture in this task.

ECCV Conference

PhD Cygert emphasizes the growing presence of Polish scientists at leading AI conferences such as ECCV. In the previous edition of this conference, there was only one article prepared by scientists from Polish institutions. In this year's edition, there are already at least six such publications. This is still a small number compared to other countries, but this trend shows that Poland is heading in the right direction.

The European Conference on Computer Vision in Milan took place from 29 September to 4 October. It is one of the most important research conferences devoted to machine learning and artificial intelligence. It takes place every two years and brings together representatives from the international world of science and industry.