Detail of publication

Abstract

The recent trend of integrating smart electronic devices into human every-day life calls for new methods of making the software both capable of performing high accuracies and meeting hardware limitations. This so called ”smartness” is often supported by sophisticated machine learning models, being developed on powerful computing machines and usually using a huge amount of data, which makes them robust and recently even surpassing human skills in a variety of cognitive tasks. The next step for a practical use, however, is to take the trained models and run them on low-cost devices, where the resources are constrained in terms of computing power and memory size. This work presents a general method for a massive reduction of parameters (80-90%) of trained (DNN or LSTM) models by removing redundant synapses. The pruning process is performed with no influence on the classification accuracy and the resulting model is minimized in terms of the number of needed parameters, which makes it suitable for on-board applications. Motivated by the recent call for keyword spotting (KWS) systems and inspired by (Chen at al. (2015)), the pruning method is shown on an LSTM-based KWS model deployed on RaspberryPi. The goal is to make the constrained low-cost device capable of spotting a keyword in real-time, with no delay, by taking advantage of the invented minimization method.