Machine Learning at the ICCUB

23 Oct 2023, 09:00 → 24 Oct 2023, 17:00 Europe/Madrid

Sala de Graus Eduard Fontserè (UB Physics Faculty, Sala de Graus Eduard Fontserè)

Tomas Andrade

Description:

Machine Learning (ML) is a powerful tool that has accelerated discoveries across numerous scientific fields. The flexibility of ML allows for the meaningful cross-fertilization of ideas among disciplines, making it an even more valuable asset. 

At the ICCUB, there are several on-going activities exploiting ML for physics and astrophysics applications. This workshop aims at facilitating the developments of on-going and nascent work at the ICCUB by showcasing the ML tools that are currently being developed. By bridging the gap between different groups, we want to take a global look at the uses of ML; identify common avenues of exploitation; and stimulate cross-fertilisation to consolidate the use of ML tools at the ICCUB and foster new potential collaborations.

Our in-person workshop will consist of a series of invited plenary overviews as well as contributed talks. We invite ML users and developers at the ICCUB to present their work, focusing on concrete applications of ML tools. To this end, we encourage you to submit an abstract to help us best accommodate your talk. We specifically aim to support the participation of early career researchers.

Scientific Organizing Committee

Tomas Andrade (ICCUB)

Teresa Antoja (ICCUB)

Arnau Rios (ICCUB)

Juan Trenado (ICCUB)

 

 

 

Acknowledgements

This event is part of the grant CEX2019-000918-M funded by MCIN/AEI/10.13039/501100011033.

 

 

 

 

 

 

 

 

Formulari_dinscripcio._Informacio_detallada_en_Machine_Learning.docx

Monday 23 October

1 Registration

2 A four philosophers' introduction to machine learning (Invited talk)

AI is everywhere and suddenly everybody is an expert on machine learning. In this talk, I will try to give a gentle introduction the main topics in the design of machine learning algorithms and what guarantees we have they will work. The relationship between ML and empirical science method will be highlighted and some applications to the Physics domain discussed.

Speaker: Oriol Pujol (U Barcelona (Math & CS))

coffee break

3 Neural Network classifier for the generation of clean Magellanic Cloud samples

The proximity of the Magellanic Clouds (MCs) to the Milky Way (MW) makes them a perfect laboratory for testing methodologies and models designed for the study of external galaxies using Gaia (ESA) data. To do so, we need to separate in the Gaia data the MCs stars from the foreground MW stars, in order to obtain “clean” MC samples. This is achieved through the design and training of a neural network classifier; the algorithm uses as much of the Gaia DR3 types of data as possible and its results have been validated through comparison with independently classified data (Jiménez-Arranz+23a,b).

Speaker: Óscar Jiménez Arranz

23.10.23 - Neural Network classifier for the generation of clean Magellanic Cloud samples [ICCUB Machine Learning workshop]-1.pdf

4 Anomaly detection for non-gaussian transient noise characterization in LIGO data using auxiliary channel information

Gravitational-wave interferometers are able to detect a change in distance of ∼ 1/10,000th the size of a proton. Such sensitivity leads to large rates of appearance of non-gaussian transient noise bursts in the main output of the detectors (the strain), also known as glitches. These glitches come in a wide range of frequency- amplitude-time morphologies and have unknown environmental and instrumental causes, hindering searches for gravitational wave transients. Current Machine Learning approaches for studying glitches use their strain morphologies to classify them but do not consider relevant information provided by auxiliary channels that are distributed through the detectors and monitor their state. In this proof-of-concept work, an unsupervised approach is taken by using the auxiliary channel data encoded in a fractal dimension measure to train an autoencoder with circular convolutions in an anomaly detection fashion. Basing the analysis on the autoencoder’s embedded space, this methodology uncovers unknown glitch morphologies, overlaps between different glitches, and misclassifications by the current state of the art.

Speaker: Paloma Laguarta

Laguarta_ML_workshop_2023ICCUB.pptx

Lunch

5 Machine Learning Techniques in Strong Gravity (Invited talk)

In this talk we will review some recent uses of machine learning techniques to perform calculations in strong gravity. These will include physics-informed neural networks (PINNs) for the solution of differential equations, and generative models such as GANs.

Speaker: Raimon Luna (U Valencia)

6 Machine learning in the context of large stellar surveys

The Gaia mission as well as large-scale ground-based spectroscopic surveys are collecting complex data for millions (even billions) of stars. Within the Gaia group we are therefore been using more and more machine-learning methods to cope with the amount of data. In this talk I will present some examples from recent publications in which we have successfully used supervised (typically regression) and unsupervised (typically dimensionality reduction and clustering) methods in the context of stellar and Galactic astrophysics.

Speaker: Friedrich Anders (Universitat de Barcelona)

fanders_iccub_ml2023.pdf

7 Symmetries in Neural Quantum States

Neural Quantum States are at the basis of a new ab-initio method especially designed to tackle the quantum many-body problem. These combine the variational method with neural networks, a flagship tool of modern Machine Learning. Neural Quantum States have been successfully used in spin, electronic and nuclear many-body systems. Neural networks can provide an unbiased approximation of complex wave functions, and so far the growth of network parameters with particle number has been found to be polynomial. One expects the growth to be further mitigated by restricting ansätze to the manifold of states that respect physical symmetries. To this end, starting from the most general way to make a neural network equivariant to a certain symmetry group, we design a many-body neural network ansatz which respects the fermionic particle-exchange symmetry. Previous fermionic ansätze are a specific case of this general approach, which we develop formally, foreseeing the need to develop relevant nuclear symmetries, like spin and isospin, into a Neural Quantum State. I will discuss how this approach will be exploited in future nuclear physics simulations and show some initial results on test systems.

Speaker: Javi Rozalén Sarmiento (Universitat de Barcelona)

Rozalén - ICCUB ML Workshop.pptx

Tuesday 24 October

8 Transforming Healthcare: Machine Learning and Computer Vision at the Forefront (Invited Talk)

Speaker: Prof. Piella Fenoy (Universitat Pompeu Fabra)

Piella_ICCUB_23.pptx

coffee break

9 Machine learning in LHCb

Machine learning techniques have a variety of use cases within the LHCb experiment. They are an essential ingredient to achieve the ultimate performance in event reconstruction and high precision in physics output.
 This talk will give an insight to the use of ML algorithms in online event selections performed by the LHCb trigger system, offline data analyses of physics measurements
, as well as to track and electron reconstruction and particle identification algorithms. Finally, the ML use in the LHCb simulation framework and the experience of the UB LHCb group with machine learning in LHCb will be discussed.

Speaker: Lukas Calefice (Universitat de Barcelona/ICCUB)

MachineLearningatLHCb.pdf

10 Quasar classification and resdhift estimation

I will present SQUEzE, a machine-learning-based quasar classifier that assigns redshifts to the classified objects. SQUEzE was originally designed to work with SDSS (DESI and WEAVE too) spectra but is highly flexible and can also cope with photometric data from multi-narrow-band photometric surveys (e.g. J-PAS). It follows the human visual inspection procedure and classifies the objects using random forest classifiers based on the found emission line peaks. The redshifts are automatically extracted from the position of the peaks.

Speaker: Ignasi Pérez (ICCUB)

20231024_SQUEzE_at_machine_learning_ICCUB.pdf

Lunch

11 Data Science Approaches to Physics and Complex Systems (Invited talk)

As physicists, we know that experiments can produce massive amounts of data. However, nowadays, we collect information from many different sources. We store different kinds of data points when we run simulations. Experiments do save data beyond what they are designed to provide (for example, metadata or instrument data). And, in complex system science, data comes from everywhere, from social networks to climatological data. Data science approaches can help generate new opportunities and exploit the vast amount of data the research produces. In this seminar, I will present some examples that go from how machine learning explanations can be used to better understand simulations to an augmented reality app that allows us to experiment and explore the human role in segregation.

Speaker: Dimitri Marinelli (UBICS)

12 Real vs. bogus classifier in time-domain surveys

Time-domain surveys are designed to study astrophysical transient phenomena appearing in the night sky. The improvements in instrumentation and data analysis are allowing the new generation of surveys to discover several thousand (and soon to be millions) of events per night. However, some of such discoveries are associated with spurious detections related to spikes from bright stars, parasitic optical reflections, cosmic rays, or defects in the detectors, just to name a few. These spurious or "bogus" detections complicate the arduous task of finding interesting real events in such a big "stream" of events. To help human scientists, several time-domain surveys have implemented Machine Learning approaches designed to automatically classify the detections as "real" or "bogus". I have contributed to the creation of real-bogus classifiers for two different time-domain surveys. In my talk, I will describe the challenge, the main algorithms used by each survey, and the main lessons learned from our efforts to compile a meaningful training set.

Speaker: Dr Nadia Blagorodnova (Univeristat de Barcelona)

13 Search for lensed gravitational-wave events with machine learning

The growing significance of Gravitational Wave Astrophysics puts in evidence the need of techniques capable of effectively and reliably analyzing all the collected data. Furthermore, the search for lensing signatures within gravitational-wave signals is a challenging task that holds the potential to uncover fresh insights into fundamental physics, astrophysics, and cosmology. In this context, we train a set of different neural networks with various types of Gravitational Wave data and we compare their performance when classifying the data in three groups: plain noise, unlensed signal and lensed signal

Speaker: Mr Roberto Bada Nerín (UPC, Barcelona)

Bada-Roberto_20231024_ML_GW-Lensing.pdf