Homanga Bharadhwaj

homanga at cs dot toronto dot edu

Hi! I am a second year graduate student in the Department of Computer Science at the University of Toronto , supervised by Florian Shkurti and Animesh Garg, and am closely collaborating with Sergey Levine at UC Berkeley. I am also a student researcher at the Vector Institute. My studies are supported by the Vector AI Scholarship and scholarships from the Department of Computer Science . I am engaged in the quest for Artificial General Intelligence. Although this quest has kept me fully occupied for the past three years, I also paint, and have a Bachelor of Arts degree in Fine Arts. Some of my paintings can be found here.

I am fortunate to have been advised by fantastic researchers over the course of my undergraduate studies. Most recently, I worked with Shin-ichi Maeda at Preferred Networks Inc. Tokyo on transferring policies to different dynamics configurations at test-time. Previously, I worked under the guidance of Yoshua Bengio and Liam Paull at Mila, Montreal during the summer of 2018. My work at Mila focused on the problem of sim-to-real transfer of deep learning based planning algorithms in robotic navigation. Prior to this, I spent a wonderful winter at NUS, Singapore in Brian Lim's lab. I continued collaborating remotely and successfully completed two projects on Recommendation Systems and Multi-Task Learning respectively. In IIT Kanpur, I have worked with Nisheeth Srivastava on Human-Computer Interaction and with Piyush Rai on Zero-Shot Learning and Program Correction.

I am currently a research scientist intern at Nvidia Research with the AI Algorithms team.

Email  /  Github  /  Google Scholar  /  Twitter  /  Name

• [Jan, 21] Excited to start a research internship at Nvidia Research for Winter, 2021
• [Jan, 21] Our papers on safety and hierarchy got accepted at ICLR, 2021
• [Dec, 20] Our papers on deep ensembles and adversarial learning got accepted at AAAI, 2021
• [Mar, 20] Excited to be visiting Sergey Levine's lab at Berekely (virtually!) during the summer+fall of 2020
• [Mar, 20] Our paper on gradient-based planning got accepted at L4DC 2020
• [Jan, 20] Our paper on policy transfer in RL got accepted at ICRA 2020
• [Oct, 19] Our paper on Zero-Shot Learning got accepted at WACV 2020
• [Aug, 19] Thrilled to be a recipient of the Vector AI Scholarship
• [Mar, 19] Our paper on exploration/curiosity in the web got accepted at ACM WebSci 2019
• [Jan, 19] I will be working at Preferred Networks Tokyo as a research intern from July to September 2019
• [Jan, 19] I am attending ICRA 2019 to present our paper on sim-to-real transfer

If you have any questions / want to collaborate, feel free to send me an email! I am always excited to learn more by talking with people.

I'm interested in deep reinforcement learning, robotics, machine learning, computer vision and explainable AI. Much of my current research is about exploration in the context of robotic policy learning. Previousy, I focused on developing efficient planning algorithms for robotic navigation/locomotion and learning transferable representations. Even before that I was involved in designing deep learning (and meta-learning) based algorithms, primarily for recommender systems and trying to make those algorithms explainable. I have explored a wide breadth of research topics around AI during my undergraduate studies, which have led to multiple primary author publications. Some of my papers are mentioned below.

I recently gave an invited talk at the Talking Robotics seminar series. link

Our research on safe exploration was covered by VentureBeat.

In an analysis of ICLR 2020 and 2019 papers, we find positive correlation between releasing preprints on arXiv and acceptance rates of papers by well-known authors. For well known authors, acceptance rates for papers with arxiv preprint are higher than those without preprints released during review.

Combining online planning of high level skills with an amortized low level policy can improve sample-efficiency of model-based RL for solving complex tasks, and transferring across tasks with similar dynamics.

Introducing skip connections in the policy and Q function neural networks can improve sample efficiency of reinforcement learning algorithms across different continuous control environments

Training a critic to make conservative safety estimates by over-estimating how unsafe a particular state is, can help significantly minimize the number of catastrophic failures in constrained RL

Updating the top action sequences identified by CEM through a few gradient steps helps improve sample efficiency and performance of planning in Model-based RL

Task-conditioned hypernetworks can be used to continually adapt to varying environment dynamics, with a limited replay buffer in lifelong robot learning

Keeping track of the currently reachable frontier of states, and executing a deterministic policy to reach the frontier followed by a stochastic policy beyond, can help facilitate principled exploration in RL

Training dynamics conditioned policies on dynamics randomized environments and estimating dynamics parameters from off-policy data can help achieve zero-shot adaptation in an unseen test environment

Adversarial domain adaptation can be used for training a gradient descent based planner in simulation and transferrring the learned model to a real navigation environment.

Explicitly maximizing diversity in ensembles through adversarial learning helps improve generalization, transfer, and uncertainty estimation

Appropriately augmenting training with multiple complimentary tasks can improve generalization in Deep Metric Learning.

Adversarially learned inference can be generalized to incorporate multiple layers of feedback through reconstructions, self-supervision, and learned knowledge.

A Generative Framework for Zero Shot Learning with Adversarial Domain Adaptation
Varun Khare, Divyat Mahajan, Homanga Bharadhwaj, VK Verma, Piyush Rai
WACV, 2020 paper code  

Adversarial Domain Adaptation appropriately incorporated in a Generative Zero Shot Learning model can help minimize domain shift and significantly enhance generalization on the unseen test classes

Recurrent Neural Network based Generative Adversarial Networks can learn to effectively model the latent preference trends of users in time-series recommendation.

A Hierarchical Multi-Task Learning Framework for Healthy Drink Recognition
Homin Park, Homanga Bharadhwaj, Brian Y. Lim
IJCNN, 2019  

A Hierarchical Multi-Task Learning model can leverage several auxiliary tasks like detection of container properties to overcome the limitation of insufficient visual cues for drinks, while predicting the healthiness of drinks from their images

Passive website recommendations embedded in the new tab displays of browsers (that recommend based on frecency) inhibit peoples' propensity to visit diverse information sources on the internet

A Meta-Learning strategy can be used to develop a recommendation model that performs resonably good enough for a wide range of users and that can be cost-effectively updated during test time for a specific user

An optimal features' classifier developed using evolutinary heuristics can be used for real time fault detection and identification

Layer-wise relevance propagation can be used for explaaining the predictions of a convolutional neural network based recommendation model

A neighborhood style explanation scheme can be used as an auxiliary mechanism for interpreting the predictions of a Recurrent Neural Network based temporal recommendation model

Phase matching in Spontaneous Parametric Down Conversion
Suman Karan, Shaurya Aarav, Homanga Bharadhwaj, Lavanya Taneja, Girish Kulkarni, Anand K Jha
Journal of Optics (Accepted 2020)

Spontaneous Parametric Down Conversion is used to generate entangled photon pairs. SPDC can be studies through the lens of Wave Optics by making some simplifying theoretical assumptions without compromising on empirical results. Also, a simulation for SPDC can be conveniently designed, given the assumptions.

• Franka Panda control environments through ROS and Gym
• L4DC, 2020 paper on Cross-Entropy and Gradient-based optimization
• Universal adversarial attack on Dilated Residual Networks
• Code for the web extension used in my paper titled "New tab page recommendations cause a strong suppression of exploratory web browsing behaviors"
• A stemmer for the Bengali language

Teaching Assitant (TA),
Introduction to Mobile Robotics (CSC477), Fall 2020

Teaching Assitant (TA),
Algorithmic Intelligence in Robotics (CSC375), Fall 2020

Teaching Assitant (TA),
Computational Cognitive Science (CS786), Winter 2019

Course Project Mentor,
Topics in Probabilistic Modeling and Inference (CS698), Winter 2019

Course Project Mentor,
Introduction to Machine Learning (CS771), Autumn 2018

Student Volunteer at ACM RecSys 2018, Vancouver, BC

Student Volunteer at ACM SIGIR 2018, Ann Arbor, Michigan

Reviewer for Computers in Human Behavior , ICRA 2019 , ICRA 2020, IROS 2020, CoRL 2020, ACM ToCHI

I love his website design.

Miscellaneous stuffs - Co-authors  /  Fav books  /  My paintings  /  Travel  /  Love  /  Name  /