✖
Soumya Sarkar

Tap on the card to flip

Soumya Sarkar

https://cpdl1997.github.io/

Soumya Sarkar /Show-Mow Shaw-r-Car/ 🎧

I am currently a PhD at the Visual Information and Processing (VIP) Lab affiliated with the Department of Electronics and Communication Engineering (ECE) at the Indian Institute of Science (IISc) Bangalore under the guidance of Dr. Rajiv Soundararajan.

I am a Master of Technology (M.Tech) graduate in Computer Science and Engineering from IIT Ropar, India, with research publications in reputed venues. I was most recently associated with the Game Theory and Machine Learning Lab under Dr. Shweta Jain at IIT Ropar. Passionate about Machine Learning (ML) and its applications in solving real-world problems, I hold an interes in both theoretical and applied ML. My research interests include, but are not limited to -

Computer Vision: Quality Assessment, 3D Novel View Synthesis
Federated Learning (FL): Client Selection Problem, Heterogenous Models
Trustworthy Artificial Intelligence (AI): Fairness, Explainability
Causal Artificial Intelligence (AI): Causal Inference, Causal Discovery
Deep Learning: Knowledge Distillation, Single class Classification

Updates

I have joined IISc Bangalore as a PhD scholar in the ECE Department!
July 2025
My paper "Cooperative SGD with Dynamic Mixing Matrices" has been accepted at ECAI 2025!
July 2025
I achieved a 8.5 band (CEFR Level C2) in IELTS Academic. You can view the certificate here.
November 2024
I am currently looking for PhD opportunities in federated learning, trustworthy ML, causal AI, or deep learning.
November 2024
I joined IIT Ropar as a JRF under Dr. Shweta Jain in the GTML Lab
August 2024

Publications

CGFLasso- Combating Multicollinearity using Domain Knowledge
CODS COMAD 2024 (IIIT Bangalore, Bangalore, India)
Soumya Sarkar, Nitin Singhal, Shweta Jain, Shashi Shekhar Jha
[PDF] | [CODE] | [SLIDES] | [POSTER]
Abstract: Multicollinearity, or near-linear dependencies amongst features, is a statistical phenomenon wherein multiple predictor variables are severely correlated in a multiple regression model. This results in inaccurate parameter estimates, decreased power, low confidence on regressor coefficients and incorrect feature selection in domains such as biological sciences or finance. This paper aims to tackle this issue by constructing a contextual graph based on available domain knowledge. We propose a novel regularization approach called Contextual Graph-guided Fused Lasso (CGFlasso), based on GFLasso regularization to show how prior domain knowledge can improve the current state-of-the-art. We show that our method significantly reduces standard error across multiple datasets. ... See More
What Did the Chicken Say?- A Multi-class Classification Method on Chicken Vocalisations
ICSTA 2023 (Kolkata, India)
Kumar Mangalam, Soumya Sarkar, Yakul Dogra, Mukesh Saini, Neeraj Goel
[PDF] | [CODE] | [SLIDES]
Abstract: Poultry is one of the largest industries in India, with a market size of USD 23.07 billion in 2022. It is imperative that the health conditions of the poultry chickens are monitored to ensure optimum output. However, due to the size of poultry farms it is a difficult task to keep track of individual chickens and ensure healthy conditions. Our current proposal aims to target this very problem statement using chicken vocalizations. We use various machine learning models to process audio signals of chicken vocalisations and try to find the best implementation combinations to get the best classification results. We find that our method gives 92.23% accuracy and shows potential in real-life applications. ... See More

‎