Forecasting multivariate time series with attention mechanism and unsupervised learning
Time: 14:00 -- Location: LRI, 445
summary: In the realm of newborn healthcare, identifying neurological pathologies has traditionally relied on the expertise of medical professionals, who perform visual assessments. However, due to the limited number of such experts available, there is an urgent need to develop a pre-diagnostic tool capable of early detection of abnormal neurological behaviors. The primary objective of our research project is to create a computational tool that leverages unsupervised learning methods to detect abnormal motor behaviors in newborns. In our previous work, we employed pose estimation models to extract numerical skeletons from 2D videos captured using conventional cameras. Our current work involves utilizing attention mechanisms for forecasting multivariate time series data. We aim to train a model on adult motricities before fine-tuning it to predict normal newborn motor behaviors and, in doing so, highlight deviations that may indicate abnormal neurological conditions.