Grants worth over PLN 2 million for researchers from the ELLIS Unit Warsaw

Bartosz Zieliński, Łukasz Kuciński and Bartłomiej Twardowski from IDEAS NCBR who are members of the ELLIS Unit Warsaw as well, received grants for a total of PLN 2,230,593 for their research projects from the Polish National Science Centre (NCN).

Bartosz Zieliński received PLN 990,600 from the SONATA BIS 13 call for the project ‘Sustainable computer vision for autonomous machines.’ This project will adapt machine learning models to new data types and create architectures for limited computing resources, prioritizing efficiency and carbon footprint reduction. The developed solutions could be used in drones to protect national parks, including preventing animal poaching. They will enable fast, efficient monitoring of large remote areas, allowing for the tracking of animal movements and early detection of forest fires.

Łukasz Kuciński received PLN 624,640 from the OPUS 26 call for the project ‘Alignment of large language models via debate and reinforcement learning,’ which aims to ensure large language models like GPT align with human values through debate-based training. Instead of using traditional methods, models are trained to argue and defend their points, engage with opposing views, and persuade judges of their stance. Employing a ‘teacher-student’ framework means AI ‘students’ will participate in debates overseen by a ‘teacher’. Outcomes will include a debate-based training protocol, public dataset of debate transcripts, trained model weights and source code, enhancing AI’s accuracy and decision-making.

Bartłomiej Twardowski received PLN 615,353 from the SONATA 19 call for the ‘MLAdapt – self-adapting machine learning’ project. The project focuses on challenges of deep learning in computer vision, aiming to create algorithms that enable continuous learning and adaptation, e.g. allowing autonomous vehicles to learn to navigate obstacles and move in difficult terrain. The models will use information available during deployment to improve their capabilities, similar to learning in humans. This approach promises applications in fields such as medicine, automotive industry, and robotics.