Research Highlights

Some recent exciting news from our research group!!
Our paper "On Translation and Reconstruction Guarantees of the Cycle-Consistent Generative Adversarial Networks" has just been accepted accepted to the NeurIPS 2022.

Our paper "GridShift: A Faster Mode-seeking Algorithm for Image Segmentation and Object Tracking" has just been accepted to the IEEE CVPR 2022 conference.

Both of our papers submitted to the NeurIPS 2021 (Thirty-fifth Conference on Neural Information Processing Systems) conference got acceptance for Spotight presentation!!

Congratuations to Swagatam Das, Anish, Saptarshi, and Debolina.

Our Research

Our contributions as part of the Machine Learning Research Group (MLRG) can be broadly categorized into theoretical analysis and applications of Machine Learning. Our research articles regularly find their rightful places in top-tier venues such as AAAI, IEEE TPAMI, IEEE TNNLS, ICCV, NeurIPS, CVPR, AISTATS, etc. The impact of our publications is further evident by the 2000+ times they are cited over recent years.

Primary Research Interests

Building Theoretical Backbone of Machine Learning Algorithms:

Following the fabled quote “There is nothing more practical than a good theory” by Lewin, we at MLRG are dedicated to investing our statistical knowledge in building the theoretical foundations of classical and emerging machine learning algorithms.

Theoretical Analysis of Clustering Algorithms:
“Clustering” algorithms attempt to find natural groups present among individuals. Even though numerous algorithms have been conceived since the days of Lloyd (1957), a significant void still marks the absence of corresponding theoretical analysis. MLRG is one of the first to employ the recent advances in learning theory to derive exact finite sample concentration results in full generality. Moreover, MLRG is invested in making clustering algorithms more efficient, fast, and interpretable. First recognized by Tukey in the early ’60s, “outliers” have consistently inspired the genesis of ‘robust’ models. Our recent paper (selected as a Spotlight submission at NeurIPS 2021) proposes a cohesive robust framework for center-based clustering under a general class of dissimilarity measures.
Deep Generative Algorithms:
In only 8 years, Deep Generative models found wide applicability and popularity in industry and academia for their commendable performance in applications like mage enhancing and processing, data augmentation, image-to-text or text-to-image translation, image-to-image visual style transfer, etc. Our Statisticians at MLRG accept their responsibility to provide theoretical foundations for deep generative models such as Generative Adversarial Networks and Wasserstein Autoencoder (WA). We are proud to point out that our regression guarantee of WAs got selected as a Spotlight submission at NeurIPS 2021.

Machine Learning Algorithms in Applications of Importance:

We at MLRG believe in applying our expertise in Machine Learning to regularly develop applications that positively impact the livelihood of society and provide artificially intelligent solutions to nationally important problems.

Addressing the Curse of Class Imbalance:
In supervised learning tasks, representatives from rare and important classes are often outnumbered. This phenomenon formally termed ‘Class Imbalance”, leads a supervised learner to be biased towards the classes with a higher number of observations. Unfortunately, class imbalance directly impacts countless lives by regularly plaguing diverse applications, such as rare disease identification, fraud detection in financial transactions, mitigating social bias in fair learners, etc. MLRG has designed techniques tailored to alleviate the effect of class imbalance in learners like k-Nearest Neighbors, Support Vector Machines, and boosted classifiers. Moreover, our work also stresses the importance of ensuring unbiased evaluation of learners in the presence of class imbalance. As put into words best by Kierkegaard, “Truth always rests with the minority…”. In search of that truth, we have proposed a generative learning-based adaptive oversampling technique that empowers minority classes, improving a deep image classifier's performance.
Machine Learning for Healthcare:
Answering the call of time, MLRG has designed an automated workflow using deep image classifiers that can detect COVID-19 cases from Chest X-Ray images. We are further progressing in this path to employ deep learning algorithms to mimic the visual signature of rare respiratory diseases for ease of detection.
Machine Learning for Social Good:
MLRG is actively investigating the scope of applying deep generative models for digitally restoring the heritage of national importance.

Our Publications

Please visit the Google Scholar page of our faculty member Swagatam Das for a detailed list of our publications. Here we are only presenting our latest works done over the past year.

You can also find implementations in MatLab or Python for many of these works in the GitHub profiles of mlrgisi, Sankha Subhra Mullick, and Shounak Datta.

Abhishek Kumar, Ajani S. Oladayo, Swagatam Das, Rammohan Mallipeddi, GridShift: A Faster Mode-seeking Algorithm for Image Segmentation and Object Tracking, IEEE/CVF Computer Vision and Pattern Recognition Conference (CVPR) 2022, Accepted for Oral Presentation, June 19 - 24, 2022, New Orleans, USA.
Anish Chakrabarty and Swagatam Das, Statistical Regeneration Guarantees of the Wasserstein Autoencoder with Latent Space Consistency, NeurIPS 2021 (Thirty-fifth Conference on Neural Information Processing Systems), Spotlight Acceptance.
Debolina Paul, Saptarshi Chakraborty, Swagatam Das, Jason Xu, Uniform Concentration Bounds toward a Unified Framework for Robust Clustering, NeurIPS 2021 (Thirty-fifth Conference on Neural Information Processing Systems), Spotlight Acceptance.
Saptarshi Chakraborty, Debolina Paul and Swagatam Das, Automated clustering of highdimensional data with a feature weighted mean shift algorithm, AAAI 2021 (35th AAAI Conference on Artificial Intelligence), pp. 6930-6938, February 2 - 9, 2021.
Saptarshi Chakraborty, Debolina Paul and Swagatam Das, t-Entropy: a new measure of uncertainty with some applications, IEEE International Symposium on Information Theory (ISIT 2021), 12-20 July 2021, Melbourne, Victoria, Australia.
Saptarshi Chakraborty, Debolina Paul, Swagatam Das, and Jason Xu, "Entropy regularized power k-means clustering", AISTATS 2020 (The 23rd International Conference on Artificial Intelligence and Statistics), Accepted 2020.
Sankha Subhra Mullick, Shounak Datta, Swagatam Das. Generative Adversarial Minority Oversampling,International Conference on Computer Vision (ICCV 2019), 2019. (arxiv link: 1903.09730).
Debolina Paul, Saptarshi Chakraborty, Swagatam Das, and Ivan Zelinka. On the non-convergence of differential evolution: some generalized adversarialconditions and a remedy. In Proceedings of the Genetic and Evolutionary Computation Conference Companion (GECCO '19), 2019, DOI:10.1145/3319619.3322007.
Sankha Subhra Mullick, Shounak Datta, Sourish Gunesh Dhekane and Swagatam Das, "Appropriateness of performance indices for imbalanced data classification: an analysis", Pattern Recognition, Elsevier, DOI: 10.1016/j.patcog.2020.107197, Accepted, 2020.
Arka Ghosh, Swagatam Das, Asit Kr. Das, and Liang Gao, "Reusing the past difference vectors in differential evolution - a simple but significant improvement", IEEE Transactions on Cybernetics, DOI: 10.1109/TCYB.2019.2921602, Accepted 2019.
Avisek Gupta, Shounak Datta and Swagatam Das, "Fuzzy clustering to identify clusters at different levels of fuzziness: an evolutionary multiobjective optimization approach", IEEE Transactions on Cybernetics, DOI: 10.1109/TCYB.2019.2907002, Accepted 2019.
Saptarshi Chakraborty and Swagatam Das, "On the strong consistency of feature weighted k-means clustering in a nearmetric space", STAT, Wiley, DOI: 10.1002/sta4.227, Accepted 2019.
Shounak Datta, Sayak Nag and Swagatam Das, "Boosting with Lexicographic Programming: Addressing Class Imbalance without Cost Tuning", IEEE Transactions on Knowledge and Data Engineering, DOI: 10.1109/TKDE.2019.2894148, Accepted, 2019.

Our Members

Faculty Member

Swagatam Das

Associate Professor
Electronics and Communication Sciences Unit
Indian Statistical Institute