TLTL: An Active Transfer Learning Method for Internet of Things Applications
Nian Wang, Tingting Li, Zhe Zhang, Li Cui · 2019
Internet of Things (IoT) devices have an increasing impact on our daily life. Part of these devices may leverage machine learning models to detect events. However, the recognition ability of a device is limited to its sensing type and the deployed position. In order to improve the system recognition ability, additional sensing devices need to be added. Nevertheless, training the computational algorithm for the newly added sensing device requires collecting and labeling large amount of data, which is both time-consuming and expensive. Many researchers are engaged to study synchronous automatic learning methods to tackle this challenge. In these approaches, learning ability of the added device can be transferred from a well-trained source device with no need to label data. However, these methods are unable to train a new model which exceeds the source model in recognition accuracy. In this paper, we propose a Talent Learner Transfer Learning method (TLTL). TLTL enables the added sensor to find new capable instances automatically, which are confusing for the old sensor but distinguishable to the added sensor, and then refines their labels through active learning. As a consequence, the added sensor can outperform the old sensor in recognition accuracy. Performance evaluations are based on three public datasets and one self-collected dataset. An average accuracy of 89.45% for the new sensor is achieved, which is 13.05% higher than the state-of-the-art method, and only 2.3% lower than the upper bound. In particular, by using TLTL method, the accuracy of the added sensor is 8% higher than the source sensor.