Hi! I'm Rahul.

I am currently a Ph.D. student in the Department of Statistics and Applied Probability at the National University of Singapore. My primary research area is Statistics and AI. Also interested in fun automations, travelling, and sketching.

Research

My research interest spans from Uncertainty Quantification, Unsupervised Learning to Deep Learning based Computer vision with application to Medical imaging, Landmark detection, and Video understanding. Here are some of my completed projects.

Uncertainty Quantification and Deep Ensembles
  Twitter    Github
										

Deep Learning methods are known to produce over-confident estimates, especially in the low data regime where the two most commonly employed techniques are Model Averaging and Data Augmentation. In our work we show that despite common belief, model averaging harms uncertainty quantification in near-calibrated/under-confident models, usually arising from data augmentation. In addition, we propose an extremely simple technique to alleviate this problem by post-processing calibration technique.

Pretrained equivariant features improve unsupervised
landmark discovery    Github
										

One of the earliest method to discover landmarks in an unsupervised fashion utilized equivariance to image deformations. We investigate one such method and find that the intermediate convolutional features suffer from poor equivaraince. To this end, we propose a two-stage training strategy that utlizes equivariant features trained from contrastive learning.

Ongoing projects

More interesting research works are on its way! Here are some of my ongoing projects.

Video Action Segmentation

It is critical for Action Segmentations in long term videos to be temporally continuous in nature. Current works on video action segmentation relies on multi-layer prediction refinement or auxilliary models to detect action boundaries and combat over-segmentation. We are working on models that do not require such additional tools and yet achieve good Edit,F1 Score as well as Frame Accuracy.

Unsupervised Landmarks and Equivariant features for Downstream Tasks

Many computer vision tasks, especially in the field of Medical Imaging, requires the model to have an overall structural understanding of the full image itself. It is infeasible to endow the models with such high-level understanding by manual supervision. Our unsupervisedly learned image features aim to provide such semantic understanding to any such downstream task.

Presentations

Introduction to Gaussian Processes

This presentation covers the very basics of Gaussian Process, its application and most of the challenges asssociated with Gassian Process models. It was presented by me during my Ph.D. seminar module at NUS.

Tutoring

I have tutored several modules in the Department of Statistics and Applied Probability at NUS.

DSA1101: Introduction to Data Science DSA3101: Data Science in Practice ST2131/MA2216: Probability ST3236: Stochastic Processes I ST5210: Multivariate Data Analysis

Industry Experience

Senior Consultant at TCG Digital

During my 2.5 years at TCG Digital, I worked on a variety of clients and projects from different fields such as Steel manufacturing, Utility distribution to Drug effectiveness testing.

Senior Business Analyst at FreeCharge

At FreeCharge, I managed customer segmentation and targeted promotions based on the history of app usage. My contributions were focused on increasing customer retention and reducing ill-utilization of promotions.

Quantitative Analyst at SPAlgo

My role at SPAlgo included devising and testing offline high-frequency trading strategies primarily based on Options, Commodities and Currency.

About Me

Contact

In case you want to get it touch, my social network profile links are on the sidebar.

Email: rahul.rahaman[at]u.nus.edu

Office: S16-04-16, 6 Science Drive 2 Singapore 117546