research

Prospective learning

Johns Hopkins

Developing theory and methods for learning under non-stationary distributions.

• 1 paper accepted at CoLLAs 2023

Inhomogeneous Lipschitzs properties of large language models

Johns Hopkins, Microsoft Research

Summary TBA.

Provably optimal domain adaptation via Fisher's Linear Discriminant

Johns Hopkins, Microsoft Research

Developed algorithms aimed at optimal and private domain adaptation for physiological state predictive systems.

The value of out-of-distribution data

Johns Hopkins

Investigated the impact out-of-distribution (OOD) data have on a supervised learning task.

• 1 paper accepted at ICML 2023
• Best short paper award at OOD-CV Workshop, ECCV 2022
• Presented at Distribution Shifts workshop at NeurIPS 2022

Kernel density networks

Johns Hopkins

Proposed an algorithm that enables a trained neural net to yield well-calibrated posteriors for both in- and out-of-distribution inputs.

• Under review in AISTATS 2023
• Presented at the NAISys 2022 workshop, Cold Spring Harbor.

Domain adaptation for in-bed human pose estimation

University of Moratuwa

• 1 Paper accepted at ICASSP 2021
• 2nd runner-up at the IEEE Video and Image Processing Cup, ICIP 2021

Deep learning-based phase unwrapping

University of Moratuwa

Developed a joint convolutional spatial quad-directional LSTM architecture for 2D phase unwrapping

• 1 Paper accepted at ICASSP 2022

Automated retinal and conjunctival vessel extraction for vascular tortuosity analysis

University of Moratuwa

Developed a novel retinal and conjunctival vessel extraction framework featuring a fully convolutional network paired with a Hessian based multi-scale vessel enhancement technique

Cost-effective active dry-contact sEMG sensor system

University of Moratuwa

Designed and developed cost-effective active dry-contact sEMG sensors and acquisition circuitry, formulated a real-time hand gesture recognition algorithm using Temporal Muscle Activation maps based on multi-channel sEMG signals, interfaced the sensors and the recognition algorithm to produce control signals to drive a bionic hand

• 1 Paper accepted at ICASSP 2020
• 1 Paper accepted at SMC 2020
• World finalists at IEEE ComSoc student competition
• Finalists at SLAAS Awards
• Finalists at SLIoT challenge

CNN-based phase unwrapping from QSM images

Center for Advanced Imaging, University of Queensland

Worked on computer vision algorithms aimed at tackling the phase unwrapping problem prevalent in MRI images derived from Quantitative Susceptibility Mapping (QSM).

Algorithms for processing and analyzing multi-electrode array signals

The Florey Institute of Neuroscience, University of Melbourne

Developed machine learning and signal processing algorithms to analyze multi-electrode array (MEA) signals acquired from in-vitro neuronal networks. These algorithms were used to study the network dynamics of various ion channel mutations responsible for genetic epilepsy

• National Finalists at Migara Ranatunga awards for best internship project