Joint Optimization of Computational Accuracy and Algorithm Parameters for Energy-Efficient Recognition Algorithm

Abstract

The need for human-machine interaction such as speech and gesture recognition has steadily grown in wearable devices. As applications become more intelligent such as facial emotion recognition, a variety of recognition algorithms has been developed and evolving. However, as the recognition algorithms become more complex, the more computation is required to perform the application in a limited battery capacity of wearable devices, which means that energy-efficiency is critical issue. In this thesis, one of the widely used recognition algorithm, artificial neural network (ANN), is selected as a target algorithm and its characteristic, inherent algorithmic fault tolerance (AFT), is adopted to improve energy-efficiency. To compute the recognition algorithm (ANN), Significant-driven iterative approximate multiplier (SDIAM) is utilized. Motivated by the fact that both an iteration of multiplication and the number of hidden nodes play key roles for a trade-off between recognition accuracy and energy consumption, these two parameters are optimized for a minimum of energy consumption of ANN, allowing acceptable recognition accuracy. The evaluation shows that the joint optimization between the iteration of multiplication and the number of hidden nodes save 70% of the energy consumption, compared with using precise computation, at the same recognition accuracy target for both handwritten and isolated spoken digit recognition. Furthermore, adopting SDIAM in training phase, the recognition accuracy is more improved, which leads to 87% and 75% lower energy consumption for handwritten and isolated spoken digit recognition. ⓒ 2016 DGIST