About the project
MLReal will be built on our ongoing project “embedded sufficient statistics” at Integreat. The outcomes of MLReal are expected to contribute to Integreat’s research themes: probabilistic modelling and uncertainty, scalable approximation of models and algorithms, and transfer learning.
Project summary
Machine learning (ML) has transformed our world with technologies such as cancer detection in medical diagnostics, generative AI that creates text, images, videos, and music, and self-driving cars. Our current ML algorithms struggle with the challenges in the real world. They are mainly designed for ideal conditions such as availability of clean and abundant data and unlimited computing resources. However, the real world is not ideal. The environment and data are constantly changing. Computing resources are limited and systems fail. MLReal aims to advance our theoretical understanding of ML and provide practical tools for ML to address imperfect conditions in the real world.
Neural networks are viewed as the artificial brains of ML systems. A central question is: How do neural networks manage to extract the most relevant information from the data? MLReal will uncover how neural networks focus only on the most relevant information within the data, what matters to make a prediction. We understand how neural networks summarize the essence of data within their layers using advanced mathematical tools. This insight opens the door for enjoying the real world. MLReal further provides efficient and robust algorithms for realistic conditions in terms of data and computing resources and addresses major challenges in important applications including marine domain and neuroscience.
Outcomes and impacts
In the long term, MLReal has the potential to provide a paradigm shift in ML theory and methods and revolutionize statistics, optimization, computer science, and engineering by reformulating the information relevance problem in addition to groundbreaking shift in the set of practical tools and working examples for highly efficient and distributed ML, and achieve one of the main goals of neuroscience beyond the field of ML.
Project team
- PI Associate Professor Ali Ramezani-Kebrya, University of Oslo, Department of Informatics
- Doctoral Research Fellow Tobias Oskar Müller, University of Oslo, Department of Informatics
- Doctoral Research Fellow Christina Runkel, University of Cambridge, Department of Applied Mathematics and Theoretical Physics
Funding
Research Council of Norway, FRIPRO programme for early career researchers
Project number: 356103