Online Reconfigurable Convolutional Neural Network for Real-Time Applications
Ahmed Shams, Mohamed Sabry, Hassan Soubra, Mohammed A.‐M. Salem · 2022
The omnipresence of real-time embedded systems enabled the smart paradigm, with smart devices being adopted by a larger portion of people worldwide. Such devices are capable of assisting people through the means of Human-Computer Interaction (HCI). Many features of these smart devices are based on Artificial Intelligence (AI) that uses neural networks (NN). Nevertheless, the computational power of embedded systems in smart devices is limited. Hence, adding more AI-powered modules becomes a challenge. Especially if the system is prone to an overload of tasks or can experience a failure that could lead to a sudden decrease in performance. To ensure that these cases are handled correctly without a total system failure, neural networks need to be able to be reconfigured quickly and effectively to cope with the available situation. This paper proposes a convolutional neural network (CNN) that can be reconfigured online at runtime to reduce its computational needs. The proposed approach changes the size of any NNs online without needing to save a whole new network to disk. This is done through offline training a network in its full size, then removing some layers and replacing them with a smaller number of layers while making sure only the new layers are being trained. After the new configuration has been trained offline, only the new layers are saved. This allows the network to be re-configurable while occupying very little disk space, simply by swapping the old layers and plugging in the new ones during runtime. This proposal has been tested on CIFAR 10 and was able to decrease the model size by up to 73.81% in only 0.18 seconds for a trade-off with accuracy.