---

About

I obtained both my PhD and MASc degrees from the Department of Electrical and Computer Engineering at the University of Toronto, Canada, in 2024 and 2018, respectively. During this period, I conducted research on distributed optimization, machine learning, and multi-user information theory under the supervision of Prof. Stark C. Draper.

I received my B.Sc. degree in 2014 from the University of Moratuwa, Sri Lanka. From 2014 to 2016 I worked with LSEG Technology (formerly known as MillenniumIT), Sri Lanka, focusing on the development of technology for Stock Market Surveillance.

Feel free to visit my Linkedin and Github pages.

I am an enthusiast of rephotography, capturing then-and-now photos. Some selected items from my collection are posted here.

---

Selected Projects + Publications

Straggler Exploitation in Distributed Computing Systems with Task Grouping  -  Presented at Allerton Conf. Sep ‘23 [Slides]

A method is proposed for exploiting work done by stragglers, nodes that randomly become slow in large compute clusters. The proposed method offers up to 30% time savings compared to a baseline.

Bounds for the Rate Distortion Region of ‘Two-Terminal Common Function Reconstruction (CFR)’ Problem  -  Presented at NASIT June ‘23 [Poster]

An achievability and a converse on the rate distortion region of the CFR problem are obtained. In CFR problem two terminals approximate a function output, and they must be identical with high probability.

Two-Terminal Source Coding With Common Sum Reconstruction (CSR)  -  Presented at ISIT June ‘22 [Paper, Slides, Arxiv]

An analysis of the CSR problem is presented. In CSR, two terminal want to reconstruct the sum of two correlated sources and they must be identical with high probability.

Compressing Gradients by Exploiting Temporal Correlation in Momentum-SGD  -  Published in JSAIT ‘21 [Paper, Arxiv]

A predictive coding-based approach for exploiting memory in consecutive momentum-SGD update vectors. A novel predictor design is proposed for systems with error-feedback.

Exploitation of Temporal Structure in Momentum-SGD for Gradient Compression  -  Presented at ISTC Aug ‘21 [Paper, Slides]

A method is proposed for exploiting the temporal structure in consecutive momentum-SGD updates for systems with and without error-feedback.

Asynchronous Delayed Optimization With Time-Varying Minibatches  -  Published in JSAIT ‘21 [Paper, Code]

TensorFlow implementation of SGD on a master-worker system with time-varying minibatch sizes, i.e., when workers complete different amounts of work due to straggling.

Decentralized Optimization with Non-Identical Sampling in Presence of Stragglers  -  Presented at ICASSP May ‘20 [Paper, Slides, Arxiv, Code]

An analysis on decentralized consensus optimization (e.g. optimization on a graph) when workers sample data from non-identical distributions and perform variable amounts of work due to stragglers (slow nodes).

Efficient Learning of Neighbor Representations for Boundary Trees and Forests  -  Presented at CISS Mar ‘19 [Paper, Slides, Arxiv, Code]

This project builds on the “Differentiable Boundary Trees” algorithm by Zoran et al. One possible implementation of their algorithm is here.

Python/NumPy Implementation of a Few Iterative Decoders for LDPC Codes   [Code]

Includes implementations of min-sum and sum-product algorithms using sparse matrices (scipy.sparse), maximum-likelihood (ML) and linear-programming (LP) decoders, and ADMM decoder.

Hosted on GitHub Pages — Theme by orderedlist