Seminar za računarstvo i primenjenu matematiku, 4. april 2023.

Naredni sastanak Seminara biće održan onlajn i uživo u utorak, 4. aprila 2023, u sali 301f Matematičkog instituta SANU sa početkom u 14.15.

Predavač: Uroš Maleš, Faculty of Electrical Engineering, University of Belgrade
 
Naslov predavanja: LEVERAGING PROBLEM UNDERSTANDING AND MACHINE LEARNING FOR BETTER SOLUTIONS: P||Cmax DIFFICULTY ESTIMATION AND TOXICITY PREDICTION CASE STUDIES
 
Apstrakt: Machine learning (ML) has emerged as a popular Artificial Intelligence (AI) technique in various scientific fields. However, the human factor remains essential in coordinating and optimizing ML models. Researchers explore the a priori knowledge that is gained through systematic use of logic and understanding of the underlying problem. We would like to present two different case studies that illustrate how ML and our understanding of underlying problems complement each other, leading to improved solutions. In the first case study we were trying to understand the difficulty two optimization algorithms have to solve P||Cmax variant of the scheduling problem. For the first algorithm (ArcFlow) we demonstrated how simple observations, made on the underlying problem, lead to a more accurate ML model. As for the second algorithm (GIST), we injected our a priori knowledge through additional feature obtained by a more sophisticated search. This resulted again in a higher predictive power of a model. The second case study focused on Toxicity Prediction using the 'Ames' dataset. We conducted a rigorous feature selection process after domain knowledge was added. That resulted in a model with significantly fewer features than in the-state-of-the-art research studies that is characterised also by the improved accuracy.  The reduced number of features in the resulting model can provide valuable insights and guide domain experts in investigating relevant factors for toxicity prediction.

The first case study is a joint work with Tatjana Jakšić-Krüger, Tatjana Davidović, Dušan Ramljak, Dragutin Ostojić, and Abhay Haridas. The collaborators for the second case study are Dušan Ramljak, Branislav Stanković, and Ranojoy Deb.

Napomena: Link za onlajn pristup:
https://miteam.mi.sanu.ac.rs/asset/YoqHWKALRkRTbK9So

Za aktivno učešće neophodna je registracija preko linka:
https://miteam.mi.sanu.ac.rs/call/wnz6oyxsQsy29LfJA/MjQ__eH607WeAL9X7IFtUI98xdQQgVkp-ljiEKPPfXr


 Ostavite vaš komentar 

Komentari(0)

Nažalost nije moguće ostaviti komentar.