Paper Title
Enabling On-Line Precision Scaling for Energy-Driven Adaptive ConvNets
Abstract
Adaptive Convolutional Deep Neural Networks (Adaptive ConvNets) can reshape their behavior to reach a better
trade-off between computational effort and prediction accuracy, which is a key feature for energy-efficient edge applications.
As a subclass, energy-driven adaptive ConvNets can self-tune their energy footprint upon request based on an external
trigger produced at the application level. This work introduces a design and optimization strategy based on the concept of
online precision scaling. The optimization was built and formulated as a multi-objective problem solved via a modified
version of the Non-Dominated Sorting Genetic Algorithm II (NSGA-II) that guarantees a fast design space exploration. The
simulation results we collected using a software-programmable neural accelerator architecture with mixed-precision
arithmetic demonstrate our approach enables ConvNets to shift over more Pareto optimal operating points, with energy
savings up to 35% for less than 3% accuracy loss.
Keywords - Deep Learning, Optimization, Energy Efficiency