SenseGAN

Abstract

Recent proliferation of Internet of Things (IoT) devices with enhanced computing and sensing capabilities has revolutionized our everyday life. The massive data from these ubiquitous devices motivate the creation of intelligent IoT systems that can collectively learn. However, labelling data for learning purposes is extremely time-consuming, which greatly hinders deployment. In this paper, we describe a semi-supervised deep learning framework, called SenseGAN, that can leverage abundant unlabelled sensing data thereby minimizing the need for labelling effort. SenseGAN jointly trains three components with an adversarial game: (i) a classifier for predicting labels of input sensing data; (ii) a generator for generating sensing data samples based on the input labels; and (iii) a discriminator for differentiating the joint data/label distribution between real samples and partially generated samples from either the classifier or the generator. The classifier and the generator try to generate fake data/labels that can fool the discriminator. The adversarial game among the three components can mutually boost their performance, which helps the classifier learn to predict correct labels with unlabelled data in return. SenseGAN can effectively handle multimodal sensing inputs and easily stabilize the adversarial training process, which helps improve the performance of the classifier. Experiments on three IoT applications demonstrate the substantial improvements in accuracy and F1 score under SenseGAN, compared with supervised counterparts trained only on the labelled portion of the data, as well as other supervised and semi-supervised baselines. For these three applications, SenseGAN requires only 10% of the originally labelled data, to attain nearly the same accuracy as a deep learning classifier trained on the fully labelled dataset.