Energy-aware Adaptive Approximate Computing for Deep Learning Applications

Application that use deep learning incur a substantial amount of energy consumption. Reducing this energy footprint is important, especially for applications such as Internet of Things (IoT) Embedded Systems (ESs), where resources are scarce. Here, we present computational self-awareness as a promising solution for intelligently adapt machine learning algorithms at runtime to reduce their energy consumption. In particular, we focus on approximation as a key enabler knob for such adaptivity. We show that the benefits of such an approach can be up to 2.5 × energy savings.