The huge amount of information that is collected nowadays combined with the high cost entailing its labeling generates a considerable number of application domains with few labelled data [Wan20]. In this scenario, the results obtained by machine learning algorithms are very reduced due to the difficult of obtaining representative models training with few labelled initial examples. In this context, semi-supervised learning [Qi20] tries to utilize all available important information both labeled and unlabeled data so that algorithms achieve their best performance. In recent years, this problem is addressed from different approaches, we focus our attention on active learning [Set09] and generative adversarial learning [Sam20]. Active learning aims to optimize the data labeling cost selecting the most representative examples that improve the algorithm performance. The challenge is to prioritize those unlabeled observations whose labels could be most informative to algorithms. Generative Adversarial Network takes a tiny portion of labeled data and a much larger amount of unlabeled data (from the same domain) and combines these sources of data to train a Deep Convolution Neural Networks (DCNN) to learn an inferred function capable of mapping a new data point to its desirable outcome.
The challenge of the few labelled data has its own characteristics in the context of temporal series, where the input data stream has an additional dimension related to its order (following a certain temporal line). Normally, the real-world problems associated to temporal series require rapid response and a broad base of prior knowledge of the most critical events [Han19]. Here, active learning and data augmentation have special potential, as can contribute to select the most significance events in the temporal series and to generate variations of them that allow the learning models to prevent critical events before they occur.
On these lines, the next advances are proposed:
- It is proposed the development of models based on active learning applied to data-intensive and weakly-labelled problems, with the aim of reducing the cost of labeling these data, as well as the response time of the learning methods by minimizing the data needed to train them.
- In the context of temporal series, the development of different techniques based on generative adversarial networks in order to increase the data base of critical and (normally) rare events it is proposed. This approach is applied as a pre-processing step able to improve the learning process of the rare events in supervised and semi-supervised learning algorithms.
- Application domains such as failures prediction in machinery is a very relevant problems in industry 4.0 and it is a problem studied by this group. This problem counts normally with a reduced number of labelled data with information about the different possible failures. It is proposed the development of new models based on active learning for anomaly detection to improve the predictions in these domains.
The research group has experience in this topic. Thus, it has carried out studies on the efficacy of active learning in different domains [Rey18, Rey18b]. Moreover, the group has addressed the imbalanced learning in support decision systems for medical diagnosis [Per20], where it is proposed a framework for data augmentation based on generative-adversarial networks.
[Han19] Z. Han, J. Zhao, Leung, H., Ma, K. F., & Wang, W. (2019). A review of deep learning models for time series prediction. IEEE Sensors Journal.
[Per20] Pérez, E., Reyes, O., & Ventura, S. (2020). Convolutional neural networks for the automatic diagnosis of melanoma: An extensive experimental study. Medical image analysis, 67, 101858.
[Qi20] Qi, G. J., & Luo, J. (2020). Small data challenges in big data era: A survey of recent progress on unsupervised and semi-supervised methods. IEEE Transactions on Pattern Analysis and Machine Intelligence.
[Sam20] Sampath, V., Maurtua, I., Martín, J. J. A., & Gutierrez, A. (2020). A Survey on Generative Adversarial Networks for imbalance problems in computer vision tasks.
[Set09] Settles, B. (2009). Active learning literature survey. University of Wisconsin-Madison Department of Computer Sciences.
[Rey18] Reyes, O., Altalhi, A. H., & Ventura, S. (2018). Statistical comparisons of active learning strategies over multiple datasets. Knowledge-Based Systems, 145, 274-288.
[Rey18b] Reyes, O., Morell, C., & Ventura, S. (2018). Effective active learning strategy for multi-label learning. Neurocomputing, 273, 494-508.
[Wan20] Wang, Y., Yao, Q., Kwok, J. T., & Ni, L. M. (2020). Generalizing from a Few Examples: A Survey on Few-shot Learning. ACM Computing Surveys, 53(3), 1–34.