Improving Energy-Efficiency of CNNs via Prediction of Reducible Convolutions for Energy-Constrained IoT Devices

Convolutional neural networks (CNNs) are widely used in various pattern recognition and computer vision applications due to their state-of-the-art performance. This superior performance comes with a tremendously high degree of computational complexity, leading to energy inefficiencies. Such inefficiencies hinder their applications on energy-constrained IoT devices, especially when it is required to perform the data processing near the source, on the edge device, for real-time requirements or security and privacy reasons. In this paper, we propose an approach to reduce a compute-intensive convolution operation by factorizing it into an inexpensive pooling operation followed by a 1 × 1 convolution operation. To identify the reducible convolutions, a lightweight, multi-class CNN-based predictor is proposed, which determines the type of the pooling operation to be applied to each activation predicted as reducible, as well. The empirical results demonstrate the great potential chances of convolution reductions using factorization by incurring a negligible storage overhead.