Empirical Study and Improvement on Deep Transfer Learning for Human Activity Recognition

Abstract

Human activity recognition (HAR) based on sensor data is a significant problem in pervasive computing. In recent years, deep learning has become the dominating approach in this field, due to its high accuracy. However, it is difficult to make accurate identification for the activities of one individual using a model trained on data from other users. The decline on the accuracy of recognition restricts activity recognition in practice. At present, there is little research on the transferring of deep learning model in this field. This is the first time as we known, an empirical study was carried out on deep transfer learning between users with unlabeled data of target. We compared several widely-used algorithms and found that Maximum Mean Discrepancy (MMD) method is most suitable for HAR. We studied the distribution of features generated from sensor data. We improved the existing method from the aspect of features distribution with center loss and get better results. The observations and insights in this study have deepened the understanding of transfer learning in the activity recognition field and provided guidance for further research.

Sensor Data Acquisition and Multimodal Sensor Fusion for Human Activity Recognition Using Deep Learning

“…Deep learning model is renowned for its feature learning ability through neural networks, which accepts a large amount of raw data for training and identify unseen data through knowledge transfer methods [31]. As a derivative of Recurrent Neural Network (RNN), LSTM network model is known to perform well on extracting signal patterns in the input feature space.…”

Literature Review on Transfer Learning for Human Activity Recognition Using Mobile and Wearable Devices with Environmental Technology

“…They hypothesis that a learning algorithm can discriminate among relevant data to learn from, and hence perform better than traditional methods when lead data are available for training. The adoptions of deep artificial neural networks is another approach which is actively utilised by the research community [19]; authors point out that some of the benefits of implementing them are given due to the automatic identification of features of a given set of data. Similarly, several architectures are flexible to be adapted to new problems and data types [20]; a characteristic that make them convenient to scale.…”

Baseline Model Training in Sensor-Based Human Activity Recognition: An Incremental Learning Approach

“…Roggen et al [13] proposed a model that consists of multiple CNN layers and LSTM layers, fine-tuned the model with different freezing layers to achieve knowledge transfer between different data sets with different distributions. Paper [14] studied distribution of sensor data of HAR and proposed an unsupervised deep transfer learning approach based on Maximum Mean Discrepancy (MMD) [33].…”

“…In the literature, a majority of published work have focused on the task of training personalized models (i.e., phase two) [4]. To this end, transfer learning based approaches have been proposed to deal with the problem of imbalanced data set, such that labeled data of a particular type of users can also help to optimize the models for other types of users [13], [14]. Moreover, incremental learning has been proposed not only to speed up the online training process, but also to provide the flexibility of learning new classes of user activities [12], [15]- [21].…”