Swami Sankaranarayanan

I am a Postdoctoral Associate at MIT working with Phillip Isola and Antonio Torralba.

Previously, I was a Research Scientist at Butterfly Network where I was part of the deep learning team led by Nathan Silberman. My responsibilities included developing and deploying interpretable machine learning models for realtime ultrasound scanning. I was an integral part of the team that developed the ML-based Auto Bladder Volume tool, among others.

I obtained my Ph.D under Prof. Rama Chellappa, for which I was awarded the ECE Distinguished Dissertation Award. In my research, I have explored different aspects of deep learning representations from the point of view of robustness and invariance. I have worked on several topics such as Semantic Segmentation, Domain Adaptation, Adversarial learning, Multi task learning and Facial recognition systems.

Broadly, my objective is to develop automated agents that mimic human intelligence not only in their ability to perceive known aspects of their environment but also embody the traits of uncertainty and causality. As machine learning algorithms become commonplace in safety critical systems such as in healthcare and law enforcement , there is a growing need for socially intelligent agents as well as informed regulation.

Email  /  Google Scholar  /  Twitter  /  Github

profile photo
Select Papers
Learning From Noisy Labels By Regularized Estimation Of Annotator Confusion
Ryutaro Tanno, Ardavan Saeedi, Swami Sankaranarayanan, Daniel Alexander, Nathan Silberman,
CVPR, 2019
arXiv / code

In addition to improving accuracy, we estimate annotator skill levels and biases.

MetaReg: Towards Domain Generalization using Meta-Regularization
Yogesh Balaji, Swami Sankaranarayanan*, Rama Chellappa,
NeurIPS, 2018
pdf / code

Learning representations that generalize to unseen domains by regularizing the structure of the neural network.

Learning from Synthetic Data (LSD): Addressing Domain Shift for Semantic Segmentation
Swami Sankaranarayanan*, Yogesh Balaji*, Arpit Jain, Ser Nam Lim, Rama Chellappa,
CVPR, 2018   (Spotlight)
arXiv / code

Using GANs to model domain shift for structured prediction problems.

Generate To Adapt (GTA): Aligning Domains using Generative Adversarial Networks
Swami Sankaranarayanan*, Yogesh Balaji*, Carlos Castillo, Rama Chellappa,
CVPR, 2018   (Spotlight)
arXiv / code

Using GANs to model domain shift for image classification problems.

Face recognition accuracy of forensic examiners, superrecognizers, and face recognition algorithms
Jonathon Phillips, Amy Yates, Ying Hu, Carina Hahn, Eilidh Noyes, Kelsey Jackson, Jacqueline Cavazos, Geraldine Jeckeln, Rajeev Ranjan, Swami Sankaranarayanan, Jun-Cheng Chen, Carlos Castillo, Rama Chellappa, David White, Alice J. O’Toole

Proceedings of National Academy of Sciences (PNAS) , 2018


Pairing automated face recognition systems with human experts on recognizing faces in the context of law enforcement.

Regularizing deep networks using efficient layerwise adversarial training
Swami Sankaranarayanan, Arpit Jain, Rama Chellappa, Ser Nam Lim,
AAAI, 2018

Exploring adversarial robustness v accuracy tradeoff by introducing regularization in the intermediate layers of a DNN.

Guided Perturbations: Self Corrective Behavior in Convolutional Neural Networks
Swami Sankaranarayanan, Arpit Jain, Ser Nam Lim,
ICCV, 2017

Adversarial and non-adversarial perturbation analysis of DNN based semantic segmentation methods.

Triplet probabilistic embedding for face verification and clustering
Swami Sankaranarayanan*, Azadeh Alavi, Carlos Castillo, Rama Chellappa,
BTAS, 2016   (NVIDIA Award for Best Paper) (IJCB Five Year Impact Award)
arXiv / third party implementation

Efficient dimensionality reduction for large scale unconstrained face recognition.

Page design borrowed from Jon Barron