Distributing a Neural Network on Axis Cameras

Sammanfattning: This document describes the methods and results of our Master’s Thesis, car- ried out at Axis Communications AB. A central problem with deep neural networks is that they contain a large num- ber of parameters and heavy computations. To cope with this, our idea was to split the network into chunks large enough that they require their own core, yet small enough to not violate our memory constraints. The goal of the thesis is to investigate whether it is feasible to distribute and run a deep neural network on a network of cameras with tight constraints such as bandwidth and memory capacity. This is done by performing experiments on existing cameras as well as Raspberry Pi’s as an assumption of how the next generation of cameras might perform. The first part of the thesis discusses how a neural network can be partitioned, and describes the problems that may occur while doing so. The second part of the thesis presents results and measurements when run on cameras and Rasp- berry Pi’s. The results and measurements are then discussed. Optimizations and bottlenecks are then described and discussed. In this part, the thesis discusses how the application benefits from hardware acceler- ation. Conclusively a few unsolved problems are identified and presented as future work.