A Survey on Light-weight Convolutional Neural Networks: Trends, Issues and Future Scope

Today with the substantial increase in the computing power of small devices and systems new challenges are emerging. For example, how to control a small handheld device which has the computing capabilities of a desktop Personal computer (PC) used five years ago. Devolving decision-making power to the device in order to make it more intelligent e.g. in the case of autonomous driving, is an interesting area. Deep learning has paved the way for this task due to its reliable decision-making capabilities which are quite popular. However for small devices there are constraints like availability of limited computation hardware, less power due to small batteries, need for real-time as well as accurate decision-making abilities, etc. In this regard, light-weight Convolutional Neural Networks (CNNs) are a valuable tool. Lightweight CNNs like MobileNets, ShuffleNets, CondenseNets, etc. are deep networks which have a much lesser number of layers and a much smaller number of parameters as compared to their larger CNN counterparts like GoogLeNet, Inception, ResNets, etc. Due to their unique advantages for small stand-alone systems, light-weight CNNs are used in these systems. In this literature survey the notable light-weight CNNs along with their architecture, design features, performance metrics, advantages, etc are discussed. The trends, issues and future scope in the area are also discussed. It is hoped that by studying this survey, the reader will engage in research in this interesting area.