Artificial Intelligence in general and deep learning specifically has experienced major breakthroughs in the last decade showing above human intelligence for complex tasks. Current deep learning technologies are however very power hungry and therefore require the use of large Graphical Processing Units (GPUs) farms in large-scale datacenters. With the recent advances in neural network hardware (neuromorphic computing design such as Neural Processing Units) we can expect more and more local neural networks being pushed to the edge of the network. However, what is lacking is exploiting the value of networking, and the Internet in particular, by connecting multiple heterogeneous learning systems together and allow more powerful learning-enabled applications to be built on top.
The goal of this project is to create a layer which is able to connect multiple heterogeneous learning systems across the Internet so that they can act as a single deep learning system performing both on-line learning and inferencing. For this, we will develop both a low overhead communication protocol and Software Defined Networking-based control layer, which can define how and when different learning systems need to be connected. Finally, we will focus on the adaptation of various learning algorithms to this connected environment so that they are able to easily transfer knowledge from one learning system to the other.