Benchmarking Remote Neural Networks

When developing Machine Learning systems, developers have multiple options where to run their models: they can for example either run them on their local machines, which are easily accessible, or on remote servers, which typically are more powerful.

This becomes particularly interesting when the development tool itself is web-based. However, sending models across the network incurrs a certain overhead, which may offset the benefit of more powerful remote hardware. In this thesis you will implement a prototype for a distributed ML system and perform benchmarks to determine when local and remote execution each make the most sense.

To this end you will:

• survey the literature on this topic,
• determine appropriate exchange formats and
• implement them measure their performance

The results of these benchmarks will be summarized and interpreted in a written thesis.