I am a graduate student at
The University of Texas at Austin
where I'm pursuing my master's in Computer Science. I work with Dr. Joydeep Ghosh at the IDEAL Lab on trustworthy AI.
I am mainly interested in tying together artificial intelligence with human-computer interaction and cognition to aid and support healthcare and other social causes. Prior to joining grad school, I had a great time at BITS Pilani, Goa where I majored in Computer Science and Physics.
Research Affiliate | Massachusetts Institute of Technology
June '20 - June '21
Working under the supervision of Prof. Pattie Maes at the Fluid Interfaces group at MIT Media Lab. We are researching AI algorithms to aid healthcare and human cognition, with a strong focus on making them fair, aware and reliable decision making systems.
Summer Analyst | Goldman Sachs
May '20 - June '20
Built a loan reconciliation app as part of the Loans Servicing Team. Wrote APIs in Java and integrated my BPMN Workflows with eTask forms. With the project succesfully moved to production, the legacy application being used was officially decommisioned.
Researcher | APPCAIR & TCS Research
January '20 - Present
Building robust and interpetable models for medical imaging under the supervision of Prof. Ashwin Srinivasan and Dr. Lovekesh Vig. Working on multiple projects involving identifying COVID-19 from Chest X-rays and lesion classification.
MITACS GRI | Western University
May '19 - July '19'
Worked as a MITACS Globalink Intern at the Nearby Galaxies group under the supervision of Prof. Pauline Barmby.
Built ImageCube, an open source image processing tool in Python to processes multi-wavelength astronomical datasets.
Summer Intern | Myra Medicine (now Medlife)
May '18 - July '18'
Worked in the intersection of business intelligence and data science supervised by Manik Singhal. Initiated a customer segmentation project and submitted a report to the marketing team for user customisation and retention.
In this paper, our focus is on constructing models to assist a clinician in the diagnosis of COVID-19 patients in situations where it is easier and cheaper to obtain X-ray data. We propose a new COVIDr dataset with important radiological annotations from a practicing radiologist. We build a deep neuro-symbolic model to diagnose COVID-19 and provide visual and textual explanations, with no significant loss in predictive accuracy compared to an end-to-end model. We find that the radiologist prefers simple representations, both visual and textual to aid in diagnosis.
Uncertainty-Aware Boosted Ensembling in Multi-Modal Settings
We propose an uncertainty-aware boosting technique for multi-modal ensembling in order to focus on the data points with higher associated uncertainty estimates, rather than the ones with higher loss values. We evaluate this method on healthcare tasks related to Dementia and Parkinson's disease which involve real-world multi-modal speech and text data, wherein our method shows an improved performance. Additional analysis suggests that introducing uncertainty-awareness into the boosted ensembles decreases the overall entropy of the system, making it more robust to heteroscedasticity in the data, as well as better calibrating each of the modalities along with high quality prediction intervals.
Disentangling Predictive Uncertainties using Deep Split Ensembles
We propose a conceptually simple non-Bayesian approach, deep split ensemble, to disentangle the predictive uncertainties using a multivariate Gaussian mixture model. The NNs are trained with clusters of input features, for uncertainty estimates per cluster. Extensive analyses using dataset shits and empirical rule highlight our inherently well-calibrated models. Our work further demonstrates its applicability in a multi-modal setting using a benchmark Alzheimer’s dataset and also shows how deep split ensembles can highlight hidden modality-specific biases
Robustness to Missing Features using Hierarchical Clustering with Split Neural Networks
In this work, we propose a simple yet effective approach that clusters similar input features together using hierarchical clustering and then trains proportionately split neural networks with a joint loss. We evaluate this approach on a series of benchmark datasets and show promising improvements even with simple imputation techniques.
CovidDiagnosis: Deep Diagnosis of COVID-19 Patients using Chest X-rays
Built a pipeline comprising of models for lung isolation, followed by classification into different disease classes. We further augment our network with symptom embeddings produced by the CheXpert network and achieve excellent results. Our visualisation maps provide trustworthy and interpretable decisions to radiologists for clinical deployment.
Studied the transfer of lesion knowledge across organs for lesion classifcation tasks. Our designed experiments on the lung and brain tumour datasets show that transfer learning using lesion-augmented models perform substantially better than models trained using random weights or lesion-agnostic(like ImageNet) transfer.
Built an open source image processing tool in Python to process multi-wavelength astronomy images. On registereing the images from different telescope to a common world coordinate system, they are convolved and re-sampled to a common pixel scale. These images are now put together to form an image cube, to help for easy analysis.
We developed a data analysis pipeline combining dataset extraction, segmentation, signal cleaning and filtration to detect the presence of sleep apnea using SVMs. On testing our approach on the MIT-Physionet dataset, we find that the low computational complexity makes it well suited for deployment on embedded devices such as the Raspberry Pi.
Online Learning Assistant with Network Community Analysis
Built a simple and scalable platform agnostic tool to aid online discussion forums using Social Network Analysis. We extract keywords from a chat, and list out the top users for this keyword and their activity histograms using a sliding window exponential ranking system. We have tested our methodology on Ubuntu IRC Logs and on our university chat for courses.
This template is a modification to Jon Barron's website. Find the source code to my website here.