The main machine learning (ML) models behind current advances in artificial intelligence are neural networks (NNs): deep and recurrent (RNNs). The latter is typically used for temporal tasks. A usual assumption when applying RNNs is that the signal or time series is uniformly sampled in time. In reality, however, data do not satisfy this condition because of multiple objective reasons. In that case, RNNs are not applicable, the time irregularity is ignored, or data are re-sampled by interpolation, in both latter cases at a loss of precision. It is similar with missing values in signals: traditional NNs like most ML methods are not suited for this type of data. However, in reality values are often missing in data because of various reasons too many to enumerate. The goal of this project is to create RNN models that could successfully function with data sampled irregularly in time, also having missing values not necessary in all inputs simultaneously; including data that have both of the features. This project is more fundamental, aimed at creating new and improving ML methods, not just applying existing ones. However, different applications will be tested with existing datasets already during the research to evaluate effectiveness of the methods.
Project funding:
KTU R&D&I Fund
Period of project implementation: 2019-04-01 - 2019-12-31
