Certain computer vision tasks using deep models need to be performed at multiple venues, namely at the cloud and the edge due to privacy requirements, platform reliability and latency requirements among others. However many state-of-the-art deep models cannot be feasibly deployed at the edge due to memory limitations and inference speed. Additionally the resources available in the environment may change, subsequently changing the computational resources.

This paper introduces TrimNet, a self-adaptive neural network designed to cater for changes in the environment by adapting the number of MACs needed to complete an inference by dynamically reducing the amount of channels used. The different trade-offs present between the accuracy against other extra-functional requirements such as inference speed and memory usage.

As part of my masters programme thesis. See the project at https://github.com/matteorapa/self-adaptive-neural-networks