An investigation into adaptive power reduction techniques for neural hardware

摘要

In light of the growing applicability of Artificial Neural Network (ANN) in the signal processing field [1] and the present thrust of the semiconductor industry towards lowpower SOCs for mobile devices [2], the power consumption of ANN hardware has become a very important implementation issue. Adaptability is a powerful and useful feature of neural networks. All current approaches for low-power ANN hardware techniques are ‘non-adaptive’ with respect to the power consumption of the network (i.e. power-reduction is not an objective of the adaptation/learning process). In the research work presented in this thesis, investigations on possible adaptive power reduction techniques have been carried out, which attempt to exploit the adaptability of neural networks in order to reduce the power consumption. Three separate approaches for such adaptive power reduction are proposed: adaptation of size, adaptation of network weights and adaptation of calculation precision. Initial case studies exhibit promising results with significantpower reductions