Opportunities and applications
11 May 2022 - 12h15-14h00
Registration mandatory - Under this link
The Data Science Competence Center (CCSD) of the University of Geneva is pleased to invite you to the eighth edition of the Data Science Seminars, exploring machine learning applications.
Machine-learning techniques are increasingly used in a variety of sciences. As a field of study in artificial intelligence, machine-learning relies on mathematical and statistical approaches to give computers the ability to "learn" from data, i.e. to improve their performance in solving tasks without being explicitly programmed for each one. Whether for analysing large datasets or for further numerical simulations, these techniques have shown that to a certain extent they can out-perform specific algorithms or human classification. Of course, like all other techniques, the use of machine-learning raises methodological challenges that should not be underestimated.
Through concrete examples drawn from their research, the speakers at this seminar will present their use of machine-learning techniques, and some of the associated limitations. These presentations will notably allow, through the example of the large hadron collider, to better understand the usefulness of machine learning techniques in the processing of large data sets. At this occasion, they will also introduce and categorise the different techniques included in the term machine-learning. The presentations of the speakers will also provide a more detailed understanding of deep convolutional models, and in particular of attention mechanisms, through concrete case studies. Finally, these presentations will offer a comparative approach of the advantages of using machine learning versus specific algorithms or human classification in High-energy astrophysics. They will also make us aware of the potential impact of using these techniques in the management of scientific knowledge.
Probing the nature of the universe with machine learning at the ATLAS experiment
Steven Schramm, Particle Physics Department.
The Large Hadron Collider is the most powerful particle accelerator ever built, allowing us to study the conditions that existed at the beginning of the universe to unprecedented precision. This accelerator discovered the highly anticipated Higgs Boson 10 years ago, providing an explanation for the origin of mass, and has turned its sights on the search for new physics and measurements of the fundamental properties of the universe.
In order to search for new physics, or measure properties of rare particles, it is necessary to sift through an enormous dataset. Processing the up to 40 million collisions per second delivered by the Large Hadron Collider, the ATLAS experiment currently records approximately ten petabytes of new data each year. On top of this large dataset, the experiment creates advanced simulations of numerous known and predicted physical processes, resulting in even larger simulated datasets.
Analysing all of this data is a massive task, and is a natural place to exploit the latest advances in machine learning. We will examine some of the ways in which machine learning techniques are being used as a key tool in the search for new physics, some of the associated limitations, and how these limitations can be mitigated.
The utility of transformers
François Fleuret, Computer Science Department
Transformer are a new class of deep neural networks using attention layers, able to dynamically modulate the importance of different parts of the signal in its computation. Such operations facilitate the learning of complex sequence processing, and those models are today the state-of-the-art solution for natural language processing.
After a short introduction to this class of models I will present three applications outside the domain of natural language processing: wind speed prediction over aircraft trajectories, image synthesis with radiance field modeling, and imitation learning in Minecraft.
Machine learning in high-energy astrophysics and its potential impact on the management of scientific knowledge
Roland Walter, Astronomy Department.
High-energy astrophysics often relies on measurements of indirect signatures which needs to be compared to simulated data to be interpreted and understood. Deep learning can find interesting results on data sets without the need for specific software. Generic deep learning algorithms often out-perform specific algorithms or human classification. I will present one application and also how machine learning could sometimes be used to help learning about experiments.
Astrophysics deals with important questions characterising the human nature and the data, and the knowledge collected on the Universe is our heritage. The United Nations is considering astrophysics data as a driver for sustainable development, promoting the use of data to foster collaboration, science, innovation and inclusion. High-energy astrophysicists pioneered and demonstrated how powerful legacy data sets can be for generating new discoveries. Artificial intelligence provides new opportunities, with a large impact on societies. People and education should be confronted with them.