Neural architecture search for sparse DenseNets with dynamic compression

Neural Architecture Search (NAS) algorithms have discovered highly novel state-of-the-art Convolutional Neural Networks (CNNs) for image classification, and are beginning to improve our understanding of CNN architectures. However, within NAS research, there are limited studies focussing on the role of skip-connections, and how the configurations of connections between layers can be optimised to improve CNN performance. Our work focusses on developing a new evolutionary NAS algorithm based on adjacency matrices to optimise skip-connection structures, creating more specialised and powerful skip-connection structures within a DenseNet-BC network than previously seen in the literature. Our work further demonstrates how simple adjacency matrices can be interpreted in a way which allows for a more dynamic variant of DenseNet-BC. The final algorithm, using this novel interpretation of adjacency matrices for architecture design and evolved on the CIFAR100 dataset, finds networks with improved performance relative to a baseline DenseNet-BC network on both the CIFAR10 and CIFAR100 datasets, being the first, to our knowledge, NAS algorithm for skip-connection optimisation to do so. Finally, skip-connection structures discovered by our algorithm are analysed, and some important skip-connection patterns are highlighted.