Research Intern, Meta AI

A key challenge in AI research is to build systems that generalize across varying data distributions. Unfortunately, the bitter lesson so far is that no proposal convincingly outperforms a simple empirical risk minimization baseline. We posit that in-context learning holds the key to better domain generalization, reformulating the problem as next-token prediction in transformers. Via extensive theory and experiments, we show that paying attention to context---unlabeled examples as they arrive---allows our algorithm to amortize performance over many out-of-distribution tasks.

Research Intern, Mila - Quebec AI Institute

In Federated learning (FL), participating clients typically each hold data from a different distribution, whereby predictive models with strong in-distribution generalization can fail catastrophically on unseen domains. We propose FL Games, a game-theoretic framework for FL for learning causal features that are invariant across clients. FL Games scales well in the number of clients, requires significantly fewer communication rounds, and is agnostic to device heterogeneity. It also achieves high out-of-distribution performance on various benchmarks.

Software Engineering Intern, Google Research

Wearables and trackers can measure several combinations of physiological markers like Heart Rate, SpO2, Stress levels, sleep durations, physical activity, etc, with varying levels of accuracy and often at different sampling frequencies. The aim of the study is to be able to learn to coach the users towards healthier lifestyle choices through smartphones. We proposed algorithms and architectures that leverage the inter-correlations among different physiological markers for time series imputation, generation, outlier and signature detection and are robust to aperiodic and erroneous data. WE further improved the learnability of both deep learning and statistical algorithms to model highly stochastic time series data.

Research Intern, Microsoft Research and Development

Capsule network is a recently introduced neural network architecture whose parameter count scales quadratically with the number of capsules per layer. We proposed a novel routing approach to improve scalability of capsule networks. We attained state-of-the-art results with capsule networks on CIFAR-10 and SVHN datasets with much fewer parameters. Our method also achieved state-of-the art performance on benchmark tasks for testing affine transformation robustness and recognizing highly overlapping digits. Our work received the Best Paper Award in MLADS-SYNAPSE 2020 and was also selected for an Oral presentation

Research Intern, Quantitative Translational Imaging in Medicine (QTIM) Lab

Model brittleness is a primary concern when deploying deep learning models in medical settings. A model which performs extremely well in one institution may plummet in performance when tested in another. This arises from inter-institution variation, such as patient demographics, as well as intra-institution variation, such as multiple scanner types. While simply combining the two datasets and re-training the model may seem like a simple solution, it is both time-consuming and fraught with data privacy limitations (for multi-institutional datasets). An alternative approach is to fine-tune the model on subsequent institutions after training it on the original institution. However, this corrodes model performance on the original dataset, a phenomenon called catastrophic forgetting . In this project, we propose a novel simple yet effective approach resulting in remarkable performance on both the original and the target domain allowing successful domain expansion while mitigating catastrophic forgetting.

Research Intern, INRIA

It is a regrettable fact that the number of road traffic accidents continues to rise, largely due to rapid urban growth and the ever-increasing density of vehicles in cities and surrounding areas.However, with emerging sensor devices and the IoT, it has become feasible to configure future vehicles with safety sensors to prevent many of these accidents. Urban traffic forecasting models generally follow either a Gaussian Mixture Model (GMM) or a Support Vector Classifier (SVC) to estimate the features of potential road accidents. Although SVC can provide good performances with less data than GMM, it incurs a higher computational cost. In this project, we propose a novel framework that combines the descriptive strength of the Gaussian Mixture Model with the high-performance classification capabilities of the Support Vector Classifier. Experimental results show that the approach compares very favorably with baseline statistical methods.

3D Rendered Ping-Pong Game

• Using multiple views of chessboard, estimated the camera caliberation matrix
• Rendered a 3D augmented reality object (ball and animals) over the chessboard
• Used video input from webcam, two visual markers as paddles reflecting the ball off the plane using laws of reflection

CovidNet: Segmenting COVID-19 abnormalities

• Developed a CT segmentation algorithm that estimates the extent of abnormality in chest CTs from COVID-19 patients
• Achieved a dice score of 0.71 on the test set with Intra-Class Correlation and Spearman coefficient as 0.99 and 0.98

My Exam Scribe

Project Presentation:    

• Built a mobile application on top of the Google Assistant using Dialogflow, Webhook and Firebase Cloud Database
• Enabled visually impaired to write exams without the use of human scribes by reading questions and storing answers