Sequence-to-sequence deep neural networks have become the state of the art for a variety of machine learning applications, ranging from neural machine translation (NMT) to speech recognition. Many mobile and Internet of Things (IoT) applications would benefit from the ability of performing sequence-to-sequence inference directly in embedded devices, thereby reducing the amount of raw data transmitted to the cloud, and obtaining benefits in terms of response latency, energy consumption and security. However, due to the high computational complexity of these models, specific optimization techniques are needed to achieve acceptable performance and energy consumption on single-core embedded processors. In this paper, we present a new optimization technique called dynamic beam search, in which the inference complexity is tuned to the difficulty of the processed input sequence at runtime. Results based on measurements on a real embedded device, and on three state-of-the-art deep learning models, show that our method is able to reduce the inference time and energy by up to 25% without loss of accuracy.
This is an open access article distributed under the Creative Commons Attribution License
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited