Curriculum Vitae

Some type of interference signals (PIM, Passive InterModulation) arriving at the receiver band frequencies complicate the decoding of useful subscriber signals. The goal of the project is to develop a machine learning model that effectively predicts interference component and has a small number of parameters for online deployment at the base station. Personal contribution at different stages includes research of classical physical and machine learning-based models for interference prediction, efficient implementation of mathematical algorithms in high-level languages, and investigation of computational complexity of different approaches.

The project is a nice example of ideas emerging from bringing two science disciplines together. On one hand, there is StarPU, a specialized library that helps computer programs run much faster by automatically splitting work between different computing nodes, like the CPU and the GPU. It takes care of the "traffic control" by using its own scheduler and thus deciding where data needs to go and which processor is best for each specific task. This library is most popular in High-Performance Computing and Numerical Linear Algebra communities, where it is used to run massive simulations, complex math problems on the most powerful supercomputers. On the other hand, there is ML applications and growing problems with the cost and availability of the most high-end GPUs. The idea was to bring these two worlds together and try to make relatively old computing clusters still relevant for the current ML demands by wise scheduling across heterogeneous computing nodes with the help of StarPU. Personal contribution at different stages includes writing mathematical code in Python, C++ and CUDA, conducting experiments on computational efficiency, and creating machine learning models on the new framework.

This was a significant early-career commercial project with a major company developing an optimization package. The challenge was that their Simplex method implementation performed slower than competitors, but we had no access to the source code. To address this, we were to develop a black-box framework that intercepted inputs and outputs of the target library, enabling us to analyze performance bottlenecks and formulate hypotheses about the underlying implementation. Based on this analysis, we proposed algorithmic improvements and independently implemented an optimized version of the solver in C++, which met the required performance specifications. This project significantly enhanced my skills in linear algebra tools, multi-level and especially low-level programming, while improving my overall code quality and programming style.