Summary
Machine learning (ML) has proven capable of tackling difficult engineering problems in image recognition and automated translation, but even more impressively in domains where traditional algorithmic approaches had struggled, such as game playing. Though the relations between ML and physics are decades old, it only recently attracted a widespread attention of scientists in many subfields of theoretical physics due to its ability to identify patterns in high-dimensional data, and to efficiently approximate complicated functional relationships. At the same time, the empirically oriented philosophy of ML is very different from that of fundamental sciences: a trained model often offers little insights into the qualitatively important aspects of the problem, how the solution was arrived at, what are the guarantees of correctness, and, crucially, how to generalize it. Bridging this conceptual gap is thus of fundamental importance, if ML is to become a powerful and controlled tool in physics research. This interdisciplinary projects aims to bring about successful development and application of ML methods resulting in qualitatively new insights in physics by following a twofold strategy. On the one hand, the performance and training of state-of-the-art ML algorithms will be improved using methods of complex and disordered systems. Specific problems targeted will include novel reinforcement learning schemes, and training of binary neural networks, with input from industrial R&D researchers. On the other, cutting edge ML techniques, particularly those with a strong underpinning in information theory, will be combined with modern computational physics methods to develop new tools for disordered systems. This is motivated by the possibility of using them to study soft materials, providing better understanding of these ubiquitous but complex systems.
Unfold all
/
Fold all
More information & hyperlinks
| Web resources: | https://cordis.europa.eu/project/id/896004 |
| Start date: | 01-09-2020 |
| End date: | 31-08-2023 |
| Total budget - Public funding: | 260 840,64 Euro - 260 840,00 Euro |
Cordis data
Original description
Machine learning (ML) has proven capable of tackling difficult engineering problems in image recognition and automated translation, but even more impressively in domains where traditional algorithmic approaches had struggled, such as game playing. Though the relations between ML and physics are decades old, it only recently attracted a widespread attention of scientists in many subfields of theoretical physics due to its ability to identify patterns in high-dimensional data, and to efficiently approximate complicated functional relationships. At the same time, the empirically oriented philosophy of ML is very different from that of fundamental sciences: a trained model often offers little insights into the qualitatively important aspects of the problem, how the solution was arrived at, what are the guarantees of correctness, and, crucially, how to generalize it. Bridging this conceptual gap is thus of fundamental importance, if ML is to become a powerful and controlled tool in physics research. This interdisciplinary projects aims to bring about successful development and application of ML methods resulting in qualitatively new insights in physics by following a twofold strategy. On the one hand, the performance and training of state-of-the-art ML algorithms will be improved using methods of complex and disordered systems. Specific problems targeted will include novel reinforcement learning schemes, and training of binary neural networks, with input from industrial R&D researchers. On the other, cutting edge ML techniques, particularly those with a strong underpinning in information theory, will be combined with modern computational physics methods to develop new tools for disordered systems. This is motivated by the possibility of using them to study soft materials, providing better understanding of these ubiquitous but complex systems.Status
TERMINATEDCall topic
MSCA-IF-2019Update Date
28-04-2024
Images
No images available.
Geographical location(s)
