Towards Deep Unsupervised Representation Learning from Accelerometer Time Series for Animal Activity Recognition

The most compelling reason to use unsupervised representation learning as a feature extraction method for effective animal activity recognition is the ability to learn from unlabeled data. Obtaining labeled data is tedious, labor-intensive, and costly, while it is much easier to obtain raw unlabeled data. In this paper, we compare three unsupervised representation learning techniques with three conventional feature extraction methods that are simple and have excellent performance. To investigate the effect of the size of both labeled and unlabeled parts of the dataset on the quality of the representations, we train the representations and classifier using various sample sizes. Furthermore, we evaluate the effect of depth in feature architectures on the performance of the representation learning techniques. All evaluations are performed on two animal datasets that are diverse in terms of species, subjects, sensor-orientations, and sensor-positions.
We demonstrate that unsupervised representation learning techniques approach and, in some cases, outperform engineered features in animal activity recognition.