Resource Efficient Machine Learning in Complex Networks

Due to their increasing size, machine learning (ML) models and data-sets are increasingly being processed in complex networks.

For example, training state-of-the-art ML models now typically requires massive parallel processing, making computations manageable but exhausting excessive communication resources.

Moreover, data is increasingly being collected and processed in IoT networks (e.g., of smartphones, home-appliances, wireless sensors) often with limited energy resources and communication over shared wireless networks with limited reliability, connectivity, and data-rates.

The success of ML in these networks is largely based on exhausting more and more communication, computation, and energy resources. This is unstainable!

The goal of this project is to advance the systematic design and theoretical foundations of resource efficient ML in complex networks.

We split the time equally between three main activities investigated in parallel: 1) communication efficient ML, how to compress algorithm information while maintaining performance? 2) ML in wireless networks, how ensure best ML performance with limited shared communication resources? 3) Energy efficient ML in IoT, how to best trade-off communication and computation resources?

Our focus is on practical algorithms with provable performance guarantees based on mathematical models verified by simulations and experiments using real IoT devices or micro-computers.