Machine Learning for Science (ML4SCI) Umbrella Organization

Mentors: Student: Muhammad Ehsan ul Haq

<p><strong>Neural AutoRegressive Flows</strong> are one of the most recent addition to the family of autoregressive flows. By using NAFs, probability density estimation in the domain of scientific exploration can yield amazing results. For example if the background data of a particular device is modelled using NAFs, it can lead to the discovery of new phenomena with very little supervision. Main goals of the project include creating basic reusable building blocks for NAFs, <em>implementing NAFs</em> on the given High Energy Physics data obtained from various experimentation devices, <em>hyperparameter tuning</em> of the NAF model for optimum performance, providing <em>API for training plus inference</em> and finally <em>documenting</em> the API and various components of the system. Final product obtained will not only be easy to use but also easy to extend.</p>

Mentors: Student: Aditya Ahuja

<p>DeepFalcon is an ultra-fast non-parametric detector simulation package. This project aims to extend DeepFalcon by adding functionality for Graph Normalizing Flows (GNFs) to it.</p> <p>While previous work on DeepFalcon has focused on other Deep Generative Models and Graph Neural Networks to simulate detector event reconstructions, we aim to use GNFs to create generative models for graph structures that can better approximate detector reconstructions.</p>

Mentors: Student: Marcos Tidball

<p>Dark matter is one of the biggest questions in current cosmology, and many different theories were created to try to explain it. One of the challenges of studying dark matter is actually finding it, as it is only noticeable due to gravitational interactions. Fortunately, recent results have shown that strong gravitational lensing can be used as a probe for studying dark matter’s substructure. Unsupervised learning algorithms are particularly interesting in this field because they allow for the identification of dark matter substructure without a prior theoretical model assumption. While the performance of these algorithms is very promising, there is still a large gap when compared to supervised learning algorithms. One promising possibility is to use domain adaptation techniques to fine tune the models trained on simulation data with real data. Thus, this project will focus on using domain adaptation to account for the differences in the modelling and available real observation data, while also improving the interface with PyAutoLens, the software used for creating the strong lens simulations.</p>

Mentors: Student: Shravan Chaudhari

<p>This project focuses on the study and implementation of Graph Neural Networks (GNNs) for low-momentum Tau Particle Identification using the CMS Open Data. The algorithm will further be deployed on CMSSW (CMS Software) inference engine for use in reconstruction algorithms in offline and high level trigger systems of the CMS Experiment. The project also aims to compare the results of GNN based algorithms with the CNN (Convolutional Neural Network) based algorithms for tau particle identification. Finally, the graph neural network algorithm will be scaled to multiple GPUs and optimised accordingly for efficient training and inference using heterogeneous computing. The final inference performance on the CMSSW inference engine will be benchmarked facilitating the end-to-end low-momentum tau particle identification.</p>

Mentors: Student: Jakub Rybak

<p>This project will seek to identify dimensionality-reduction methods that achieve a reduction in the number of features while maintaining predictive performance of the relevant models. This will be done in the context of astrophysical data, however, the problem of many features is common in many areas, including genomics, economics and imaging.</p>