Congratulations Sarit Khirirat

“Traditionally, optimization algorithms were limited by the available computing resources. But with the rise of big data and large-scale machine learning, the bottleneck has shifted from computations to communications. It is not uncommon that 80% of the run-time is devoted to communication between participating computers. In this work, we demonstrate how we can compress data by several orders of magnitude while essentially keeping the same solution accuracy. Specifically, the paper studies techniques for compensating for the information loss due to gradient compression, and provides the first theoretical support for the improved accuracy of these methods.”

In collaboration with Dr. Sindri Magnússon from Harvard University and supervision of Professor Mikael Johansson.