Witold Kościukiewicz and Mateusz Wójcik, implementation doctoral students of our Faculty, received the first two awards (Best Poster Award and Runner-Up Best Poster Award) at the ML in PL conference in Warsaw. 34 posters were submitted to the competition. Our winners prepared posters summarizing their doctoral research. Congratulations!

Witold Kościukiewicz received the 1st prize for the poster Memory Enhanced Document-level Joint Entity and Relation Extraction, and Mateusz Wójcik received the 2nd prize for the poster Neural Architecture for Online Ensemble Continual Learning. They both work at Alphamoon, preparing implementation doctorates under the supervision of Adam Gonczarek, Ph.D. Their university supervisor is Prof. Tomasz Kajdanowicz, head of the Department of Artificial Intelligence.

Witold Kościukiewicz is writing his Ph.D. thesis on "Knowledge Extraction From Unstructured Text Data." Mateusz Wójcik – on "Continuous Learning Methods in Natural Language Processing."

– We took part in this year's edition of the ML in PL conference in Warsaw with a poster session in which people interested in machine learning – students, doctoral students and researchers – could present their work. The session included 34 posters and involved a competition for the best work, which we won – says Mateusz Wójcik.

Witold Kościukiewicz's poster "Knowledge Extraction from Unstructured Text Data" was devoted to his research carried out so far on machine deep learning methods in the field of natural language processing.

– The main goal was to develop a method for extracting relationships between named-entities in text documents, detected by methods already implemented in our company. One of the research results is a method that allows the extraction of relationships with a higher efficiency than the methods existing in the literature so far, when the task is to extract the graph from the text, performing four tasks end-to- end (mentions, coreferences, named-entities and relationships) – says the winner of the 1st prize.

According to Witold Kościukiewicz, a large part of the work additionally comprises an analysis of Alphamoon's clients' requirements in the context of the relationship extraction task: what do they need, what data do they have, how their needs coincide with the research objective and the leading trends in the field of relationship extraction in the literature. The research also includes the preparation of data sets based on real customer data.

Mateusz Wójcik's poster "Neural Architecture for Online Ensemble Continual Learning" investigated the challenging problem of Class Incremental Continual Learning.

– Inspired by the work of researchers in the field of continuous learning based on team methods and the Mixture of Experts approach, we proposed an improved architecture. In our work, we extended the model with a differentiable layer approximating the result of the KNN algorithm; we proposed an innovative method of aggregating team votes and performed extensive research in various configurations of data sets and our model, in which we compared 10 methods of continuous learning – explains Mateusz Wójcik.

Mateusz Wójcik adds that the method:
- demonstrated higher accuracy compared to the reference method, using the same number of parameters,
- allowed to limit the phenomenon of forgetting,
- allowed to obtain higher results compared to the methods with a memory buffer,
- was better than baseline methods in all evaluated configurations.

The obtained results indicate that the chosen direction of continuous learning with the help of team methods can be continued and demonstrate many advantages of the method.