AETHER: Augmentation and Episodic Task Harnessing for Efficient Recognition

Abstract

Human Activity Recognition (HAR) has become a significant research area for human behavior analysis. Researches from the middle from the last decade prove that deep learning based models are suitable to identify patterns over time series data collected from smart devices (smartphones, smartwatches) and perform accurate activity recognition over a fixed set of observed activities. However, deep learning approaches face some challenges for time series data, like the lack of sufficient data to train an efficient model. Another challenge that comes with HAR is the particularities in the way that users perform the same activity or how the sensor collects the data, generating some individual conditions. This Ph.D. thesis presents a meta-learning algorithm that overcomes the individual condition limitation by providing a training strategy that facilitates the generalization across different tasks, allowing the model to adapt to unseen users, sensors, and activities. It also overcomes the labeled data scarcity limitation by proposing a data augmentation stage to increase the number of observations to be used during the meta-training. The algorithm is compared against the literature using real-world public datasets and obtains encouraging results. The algorithm is compared against the literature using real-world public datasets and obtains good results, surpassing some literature baselines by 20%. Furthermore, the trained meta-models are applied against other public datasets, allowing us to evaluate the meta-models in completely new scenarios, where the proposed algorithm was able to overcome, in some cases, the baselines by over than 40%.