Energy-Accuracy Scalable Deep Convolutional Neural Networks: A Pareto Analysis

Abstract

This work deals with the optimization of Deep Convolutional Neural Networks (ConvNets). It elaborates on the concept of Adaptive Energy-Accuracy Scaling through multi-precision arithmetic, a solution that allows ConvNets to be adapted at run-time and meet different energy budgets and accuracy constraints. The strategy is particularly suited for embedded applications made run at the "edge" on resourceconstrained platforms. After the very basics that distinguish the proposed adaptive strategy, the paper recalls the software-to-hardware vertical implementation of precision scalable arithmetic for ConvNets, then it focuses on the energy-driven per-layer precision assignment problem describing a meta-heuristic that searches for the most suited representation of both weights and activations of the neural network. The same heuristic is then used to explore the optimal trade-off providing the Pareto points in the energy-accuracy space. Experiments conducted on three different ConvNets deployed in real-life applications, i.e. Image Classification, Keyword Spotting, and Facial Expression Recognition, show adaptive ConvNets reach better energy-accuracy trade-off w.r.t. conventional static fixed-point quantization methods.

A Procedural Method to Predictively Assess Power-Quality Trade-Offs of Circuit-Level Adaptivity in IoT Systems

“…OnA allows finer control over the power/quality trade-off at run time, as demonstrated by several recent works, e.g., the work of Peluso and Calimera (2019);De Roose et al (2020);and Teo et al (2020). When a minimum required quality is more easily met in specific environments, it allows for less-power-intensive settings (Figure 4).…”

“…It includes margins according to the application or task, the current usage (the same application may require several quality levels), or the dataset (different inputs of the same application and context may require different quality levels), which can be traded off against energy in context-aware adaptivity. EQ scalability methods have already been applied to numerous designs, e.g., sensor nodes (Badami et al, 2015;Cao et al, 2017;Ieong et al, 2017;Xin et al, 2018;De Roose et al, 2020), digital circuits, (Moons and Verhelst, 2014;Rizzo, 2019), and edge Artificial Intelligence (AI) hardware Peluso and Calimera (2019). It has also been identified among key factors to enable the next-generation IoT nodes and communications nodes for 5G and 6G Mahmood et al (2020); Shafique et al (2020).…”

Energy-Quality Scalable Monocular Depth Estimation on Low-Power CPUs

“…The scaling can be performed at different granularities, e.g., per-layer or per-network. The former assigns a different arithmetic precision to each layer [21], [53], [54], the latter operates the same scaling on the whole network. The second option is preferred for general-purpose cores, such as those targeted in this work, thanks to its simple implementation.…”