profile image

Shravan S. Chaudhari

Domain Adaptation | OOD Detection | Graph Neural Networks | Computer Vision

Contact Me

About Me

Hi! I am a final year MS CS student at New York University (Courant Institute of Mathematical Sciences).

I’m interested in semi-self supervised representation learning, using it in identifying knowledge gaps in machine learning models as well as improving robustness and generalization of ML algorithms.

Previously, I have worked in using active learning and data centric machine learning to solve important problems in the biomedical and healthcare domains. Before that, I worked on using machine learning to solve problems in high energy physics, notably in CMS experiment at CERN to identify and classify the energy deposits recorded at the Large Hadron Collider after the particle collision.

Feel free to contact me for research collaborations or other engagements.

Industry Experience

Research Experience

project name

Graph Neural Network based High Energy Particles Reconstruction

Efficient and accurate reconstruction of Tau lepton decays (elemetry particle like electron) is crucial for the study of high energy particle colliders. The particle collisions recorded at the Large Hadron Collider (LHC) are analyzed by physicists all around the world. The recorded energy decays are a mix of a variety of signals and noises. Hence, it becomes necessary to correctly identify Tau signals from background noises from other particle decays. This is an important step in Particle Reconstruction in the CMS Experiment at CERN. Deep Learning has provided more than sufficient evidence of its potential in using the raw decay images from LHC directly for particle classification and identification. However, the sparse structure in Tau decays restrict the capability of Convolutional Neural Networks (CNNs) accurately classify Taus from background signals of QCD jets, W jets and TTbar jets. Moreover, Taus are more complicated than previous benchmark datasets of electron vs photon classification (1 channel decay images), Quarks vs Gluons (3 channel decay images), Boosted Top Jets (8 channel decay images), etc. in that Taus have greater number of decay image channels - 13 channels.

As part of my research fellowship under the mentorship of Dr. Sergei Gleyzer , I proposed a Graph Neural Network (GNN) based solution to leverage the sparse structure in the energy decays of the particle hits at the LHC with the help of inductive representation learning on graphs. This method outperformed several state of the art approaches like CNNs and Vision Transformers particularly for low momentum Tau particle identification.

Find out more

project name

Deep Learning for CMS Experiment at CERN

Developed novel deep convolutional neural network (DCNNs) based solutions for classification and reconstruction of high energy particles like electrons, photons, quarks, gluons and boosted top jets. Presented detailed analysis of the strengths and weaknesses of ResNet based Deep Convolutional Neural Networks (DCNNS) on various energy decay channels like electromagnetic calorimeter, hadronic calorimeter and characteristics of particles such as location of the energy deposit and transverse momentum recorded by silicon detectors at the Large Hadron Collider.

Find out more

Open Source Contributions

project name

Google Summer of Code 2021

Designed & implemented graph based modelling strategies to represent and classify low momentum particle images having 13 energy channels recorded by various silicon detectors at the Large Hadron Collider. Developed a standard framework to deploy graph neural networks (GNNs) within CERN software pipeline using C++ ONNX API and Torch-Geometric library. Key accomplishments: Benchmarked the graph neural network performance and its deployment using the newly built framework for classification of electrons, photons, quark, gluons and boosted top jets by achieving ROC AUC score of atleast 0.95 for each category.

Find out more

Google Summer of Code 2020

Created a C++ based End-to-End Software (E2E) framework to enable advanced data processing and complex analysis on the CERN database. Integrated the E2E framework with the CERN inference engine to support deployment of machine learning architectures like GNNs, CNNs, Variational Autoencoders, etc. trained with either of the 4 different frameworks: tensorflow, mxnet, onnx and pytorch.

Find out more