Multi-task Self-Supervised Learning for Human Activity Detection
Aaqib Saeed, Tanir Ozcelebi, Johan Lukkien
Eindhoven University of Technology, Netherlands
Embedded Files
Abstract
Abstract
Existing supervised methods need a massive amount of well-curated data to extract generalizable features. The task of collecting a large labeled sensory data in a real-world setting is notoriously challenging, in particular, for applications pertaining to ubiquitous computing, pervasive intelligence, and health-care. Transformation Prediction Network instead utilizes self-supervision to learn representations in an unsupervised manner without using any semantic labels; making it suitable for semi-supervised and transfer learning in the low-data regime. Moreover, it can also be used for continual multimodal learning from unlabeled sensory streams; creating an immense opportunity for unsupervised feature learning in an Internet of Things era.
Approach
Approach
The objective of our work is to learn general-purpose sensor representations using a temporal convolutional network in an unsupervised manner. To achieve this goal, we introduce a self-supervised deep network named Transformation Prediction Network (TPN) to recognize the applied deformation on the raw input signal. Importantly, the TPN is trained in a multi-task learning setting to solve multiple binary classification problems simultaneously. The network has a common trunk (an encoder) and an individual head for each task; it takes an input sequence and produces a probability of the signal being a transformed version of the original. We generate transformed versions of the signals for the self-supervised pre-training of the network. At each training iteration of the TPN model, we feed data from all tasks, and overall loss is calculated as a weighted sum of the losses of individual tasks. Once pre-training converges, we transfer the weights of the encoder to an end-task classifier (e.g., activity recognizer) for learning the supervised task.
Results
Results
Comparison against fully-supervised and unsupervised approaches
Comparison against fully-supervised and unsupervised approaches
Assessment of individual self-supervised tasks in contrast with multiple tasks
Assessment of individual self-supervised tasks in contrast with multiple tasks
Effectiveness under semi-supervised setting
Effectiveness under semi-supervised setting
Evaluating knowledge transferability
Evaluating knowledge transferability
Determining representational similarity
Determining representational similarity
SVCCA similarity between fully-supervised and self-supervised networks
Citation
Citation
@article{saeed2019multi, title={Multi-task Self-Supervised Learning for Human Activity Detection}, author={Saeed, Aaqib and Ozcelebi, Tanir and Lukkien, Johan}, journal={Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies}, volume={3}, number={2}, pages={61}, year={2019}, publisher={ACM}}Page updated
Google Sites
Report abuse