Neuroimaging Signatures of Brain Disorders: Fighting Overfitting in Predictive Models.
Edouard Duchesnay, CEA Neurospin
10 décembre 11h30, en distanciel
My presentation will summarize my research in designing machine learning models to discover brain imaging signatures of mental disorders.
I will present dimension reduction and regularization strategies to overcome the “curse of dimensionality” caused by a large number of neuroimaging measurements.
Given the limitations of sparse models to produce stable and interpretable predictive signatures, I proposed to push forward regularization by integrating spatial constraints. Evaluations on experimental data demonstrated that those constraints force the solution to adhere to biological priors, producing a more plausible interpretable predictive brain signature of clinical status.
To bridge the gap between biological processes and brain imaging, I present multivariate latent variable sparse models to investigate the genetic influence on the brain.
Machine Learning for Astronomical Data Analysis
François Lanusse, CEA Irfu
20 novembre, 11h30 14h, en distanciel
Reconnaissance automatique des détecteurs Atlas/NSW sur la plateforme de métrologie CICLAD à Saclay
Michel Mur, CEA Irfu
REPORTE 10 novembre, 11h30, CEA-Saclay, bat 713, salle Galilée
Le site de Saclay est l’un des 4 sites de production des New Small Wheels, de nouveaux détecteurs Micromegas destinés à être installés prochainement sur l’expérience Atlas au CERN. Ces détecteurs de grande taille (3 m2) sont constitués d’un assemblage de panneaux composites construits sur des marbres instrumentés dans la salle blanche CICLAD. La planéité des panneaux est mesurée en place par un portique mobile portant un capteur optique sans contact, lors des étapes intermédiaires de fabrication et lors du contrôle qualité final.
Afin d’automatiser et fiabiliser le processus, un système de reconnaissance automatique des éléments mesurés aux différentes étapes a été mis en place à titre exploratoire en 2020. Ce système s’appuie sur l’analyse d’une série d’images de l’élément en cours de mesure, qui sont soumises indépendamment pour classification à un réseau de neurones. La fusion des résultats obtenus est ensuite utilisée pour définir automatiquement les réglages machine liés à cette configuration particulière.
L’exposé décrit les choix de l’architecture de classification à catégories multiples, les campagnes de collecte d’images et l’augmentation des données, la stratégie d’apprentissage par transfert et l’analyse des résultats de fusion. Il aborde ensuite l’intégration du réseau pour inférence dans le programme de pilotage et présente les premiers résultats obtenus.
Multiwavelength classification of X-ray selected galaxy cluster candidates using convolutional neural networks
Matej Kosiba, CEA Irfu
1er septembre, 11h30, CEA-Saclay, bat 713, salle Galilée
Galaxy clusters appear as extended sources in XMM-Newton images, but not all extended sources are clusters. So, their proper classification requires visual inspection with optical images, which is a slow process with biases that are almost impossible to model. We tackle this problem with a novel approach, using convolutional neural networks (CNNs), a state-of-the-art image classification tool, for automatic classification of galaxy cluster candidates. We train the networks on combined XMM-Newton X-ray observations with their optical counterparts from the all-sky Digitized Sky Survey. Our data set originates from the X-CLASS survey sample of galaxy cluster candidates, selected by a specially developed pipeline, the XAmin, tailored for extended source detection and characterisation. Our data set contains 1 707 galaxy cluster candidates classified by experts. Additionally, we create an official Zooniverse citizen science project, The Hunt for Galaxy Clusters, to probe whether citizen volunteers could help in a challenging task of galaxy cluster visual confirmation. The project contained 1 600 galaxy cluster candidates in total of which 404 overlap with the expert's sample. The networks were trained on expert and Zooniverse data separately. The CNN test sample contains 85 spectroscopically confirmed clusters and 85 non-clusters that appear in both data sets. Our custom network achieved the best performance in the binary classification of clusters and non-clusters, acquiring accuracy of 90 %, averaged after 10 runs. The results of using CNNs on combined X-ray and optical data for galaxy cluster candidate classification are encouraging and there is a lot of potential for future usage and improvements.
Designing superconducting magnets using simple genetic algorithms
Valerio Calvelli, CEA Irfu
9 juillet 2020, 14h, CEA-Saclay, bat 141, salle André Berthelot
Designing superconducting magnets is a complex task: on one side, the different objectives and constraints of the design may results in intricate 2D and 3D windings; on the other side, the presence of ferromagnetic materials makes the magnetic field a non-linear function of the current distribution. Working with FEM models to take into account all these challenges performing parametric studies is a consolidated practice, but it can require a long time before reaching the desired result. Used since the beginning of the 80’s, Genetic Algorithms have been found as one of the easiest AI method to reduce the computation and person-time spent in designing, and their simple implementation makes them adapt to any FEM software used. Advantages and backwards of using this method will be explained in this seminar showing two cases: the iron optimization of the iron around F2D2, the flared-ends Nb3Sn dipole prototype for the Future Circular Collider at CERN, and the 3D shape optimization of the MADMAX dipole, the Max Planck experiment to discover axion-like particles.
Interpretable machine learning for CLAS12 data analysis
Noëlie Cherrier, CEA Irfu et List
7 février 2020, 11h30, CEA-Saclay, bat 714 (ICE), salle visio 1110
The Generalized Parton Distributions (GPDs) describe the correlations between the transverse position and the longitudinal momentum of the partons (i.e. quarks and gluons) inside the nucleon. They can be extracted from exclusive inelastic processes, i.e. processes with a fully characterized final state. In the Hall B of the Jefferson Laboratory, the CLAS12 collaboration probes the inner structure of the proton by colliding 11 GeV electrons into a fixed proton target. Among the exclusive inelastic processes that are produced, we focus on the Deeply Virtual Compton Scattering (DVCS) in which the collided proton emits a high-energy photon. The objective is to be able to isolate these events in CLAS12 data, and notably separate the DVCS from mimicking exclusive Pi0 production events since Pi0 decays instantly into two photons.
This talk will focus on the use of interpretable machine learning algorithms to perform event classification in CLAS12 data. Interpretable or transparent algorithms are preferred for the sake of trust or for further understanding the patterns in the data. On the contrary to black-box models such as neural networks, transparent models such as decision trees, rule bases or Generalized Additive Models (GAM) are more easily understood and validated for further physics analysis. However, the performance of these models has an increased dependency on the input variables, since their internal representation is not as complex as the one of neural networks. In order to increase the performances of such interpretable model, we will address in this talk the inclusion of prior physics knowledge by building high-level variables that are the most relevant to the task and using assumptions on their distributions.
