Radhika Dua

Hi, I am a Predoctoral Researcher at Google Research India, where I am working under the guidance of Dr. Gaurav Aggarwal. Before joining Google, I was a Master's student in Graduate School of AI at Korea Advanced Institute of Science and Technology (KAIST), where I was fortunate to have Prof. Edward Choi as my advisor. My research interest lies at the intersection of Computer Vision, Machine Learning, Healthcare, and NLP. I currently work on enhancing the reliability of deep neural networks against natural and synthetic distribution shifts.

Prior to my time at KAIST, I was a Summer Intern at Brown University, where I was fortunate to be advised by Prof. Srinath Sridhar. I also spent some time as a Visiting Researcher at IIT Hyderabad, where I was blessed to be supervised by Prof. Vineeth N Balasubramanian.

Email  /  CV  /  Google Scholar  /  Github /  Twitter

profile photo
News
Selected Publications
Please refer to Google Scholar or Publications section of CV to view all papers.
ConDor: Self-Supervised Canonicalization of 3D Pose for Partial Shapes.
Rahul Sajnani, Adrien Poulenard, Jivitesh Jain, Radhika Dua, Leonidas J. Guibas, Srinath Sridhar
CVPR 2022
Project Page / Paper / Video / Code

ConDor is a self-supervised method that learns to Canonicalize the 3D orientation and position for full and partial 3D point clouds. It can also learn to consistently co-segment object parts without any supervision.

Natural Attribute-based Shift Detection.
Jeonghoon Park*, Jimin Hong*, Radhika Dua*, Daehoon Gwak, Yixuan Li, Jaegul Cho, Edward Choi
Under Review
Paper

We define a new task, natural attributebased shift (NAS) detection, to detect the samples shifted from the training distribution by some natural attribute such as age of subjects or brightness of images. We also introduce benchmark datasets in vision, language, and medical for NAS detection.

Beyond VQA: Generating Multi-word Answer and Rationale to Visual Questions.
Radhika Dua*, Sai Srinivas Kancheti*, Vineeth N Balasubramanian
MULA Workshop, CVPR 2021
Paper / Slides / Video / Poster

We introduce a new task: ViQAR (Visual Question Answering and Reasoning), wherein a model must generate the complete answer and a rationale that seeks to justify the generated answer.

VayuAnukulani: Adaptive Memory Networks for Air Pollution Forecasting
Radhika Dua*, Divyam Madaan*, Prerana Mukherjee, Brejesh Lall
GlobalSIP, 2019
Paper / Code / slides

We present VayuAnukulani system, a novel end-to-end solution to forecast fine-grained ambient air quality information based on the historical and realtime ambient air quality and meteorological data.

Experience
Summer Intern, Brown University
Supervisor: Prof. Srinath Sridhar

Conducting research in 3D computer vision and machine learning.

Research Intern, Indian Institute of Technology, Hyderabad
Supervisor: Prof. Vineeth N Balasubramanian

Conducting research in Vision and Language applications and introduced a new task, ViQAR: Visual on Answering and Reasoning, which focuses on automatic generation of the answer, and of a rationale, given a visual query.

Research Intern, Celestini Project India
Mentor: Dr. Aakanksha Chowdhery (Google Brain and Tensorflow) and Prof. Brejesh Lall (IIT Delhi).
Sponsors: Marconi Society and Google

Developed a temporal forecasting solution based on the historical data reported by Central Pollution Control board to predict the real-time and fine-grained air quality information in five locations of Delhi.

Teaching
NPTEL: Deep Learning for Computer Vision
Instructor: Prof. Vineeth N Balasubramanian

IIT Hyderabad-CS6360: Advanced Topics in Machine Learning - Spring 2020
Instructor: Prof. Vineeth N Balasubramanian

IIT Hyderabad-CS5370: Deep Learning for Computer Vision - Fall 2019
Instructor: Prof. Vineeth N Balasubramanian

KAIST-CC500: Scientific Writing
Instructor: Prof. Mik Fanguy

KAIST-HSS583: Graduate English Presentations
Instructor: Prof. Mik Fanguy
Selected Projects
ViQAR: Visual Question Answering and Reasoning
Radhika Dua, Sai Srinivas Kancheti, Prof. Vineeth N Balasubramanian

We introduced a new task, ViQAR, in which the model generates the complete answer and rationale. We also proposed an end-to-end, attention-based encoder-decoder architecture to solve this task, and showed that our model generates strong answers and rationales through qualitative and quantitative evaluation, as well as human Turing Test.

Clair: Air Pollution Prediction
Radhika Dua, Divyam Madaan, Dr. Aakanksha Chowdhery

As a part of the Celestini Project, sponsored by Google and Marconi Society and mentored by Dr. Aakanksha Chowdhery, we developed Clair: Air pollution forecasting in Delhi, a temporal forecasting solution based on the historical data reported by Central Pollution Control board to predict the real-time and fine-grained air quality information in five locations of Delhi. We were awarded Second Prize for our novel solution and also had the opportunity of presenting our work at GlobalSIP 2019.
website / video / Financial Express / Hindustan Times / NDTV / Business Standard / First Post / India Today / Hindustan Times / code


Design and source code from Jon Barron's website.