| Grant number: | 23/02680-0 |
| Support Opportunities: | Scholarships in Brazil - Master |
| Effective date (Start): | April 01, 2023 |
| Effective date (End): | April 30, 2025 |
| Field of knowledge: | Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques |
| Principal Investigator: | Diego Furtado Silva |
| Grantee: | José Gilberto Barbosa de Medeiros Júnior |
| Host Institution: | Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil |
| Associated research grant: | 22/03176-1 - Machine learning for time series obtained in mHealth applications, AP.PNGP.PI |
| Associated scholarship(s): | 23/11775-5 - Cross-convolution for multivariate time series tasks, BE.EP.MS |
Abstract One of the limitations to the wide use of mHealth technologies is heterogeneity between devices and from one user to another. In other words, different devices may have differences in the obtained signal, which may occur, for example, due to differences in sensitivity to ambient light and artifacts due to interference from other device components. Even if subtle, these differences can cause significant damage to a model's performance. A solution to this problem would be to perform a new training data collection for the new device and train a new neural network. However, this is a very costly task, especially due to data collection and labeling. In the case of deep neural network models, training usually requires a large amount of annotated data to achieve good performance. Alternatively, it is possible to train a network with a large volume of data collected on a given device and use this network as a starting point for a model that can be used for a second device. In this case, it is enough to have a small number of representative examples collected in the second device and train the neural network for a few more epochs, a process known as refinement. This strategy is an example of transfer learning. Despite having shown good results in several domains, transfer learning, as described, is not scalable for the mHealth context. With each new device model that can be used in mHealth applications, new predictive models would have to be trained and, for that, new data collection and annotation performed. In this scenario, this work proposes the adaptation and creation of unsupervised transfer of learning techniques. In the general scheme of transfer learning, it is common to consider that the new dataset is labeled. In our proposal, for this strategy to be easily extensible to any device, this dataset has no labels. The challenge, in this case, is to create or adapt a transfer strategy that does not rely on annotations for new data. To the best of our knowledge, in addition to the fact that there are few proposals like this in the literature, none of them have been applied to time series. Thus, this activity will be able to contribute to the Machine Learning area for time series in its entirety, in addition to the specific domain of Health. | |
| News published in Agência FAPESP Newsletter about the scholarship: | |
| More itemsLess items | |
| TITULO | |
| Articles published in other media outlets ( ): | |
| More itemsLess items | |
| VEICULO: TITULO (DATA) | |
| VEICULO: TITULO (DATA) | |