Improving Models for User-Specific State Assessment: A Real-time Querying and Learning Technique
Advances in wearable physiological sensors can now potentially provide critical information about human functional states in real-time, in order to support performance, learning, and safety in various work, training and leisure-related contexts. It has been observed that models trained for human state prediction in laboratory suffer from drastic performance loss when deployed on subjects unseen by the model. In this project, we propose to use systems made of multiple models, and calibrate decisions made with these pre-trained models on new users. For instance, integration of the models will be dynamically reweighted for a new human subject using the system according to similarity and uncertainty measures. One element of novelty lies in adding the possibility to query users to improve the combination of models. It is expected that this methodology will improve performance over current models for new human subjects. This will make Thales’s human functional state prediction platform more robust and able to adapt
dynamically to new users.
Christian Gagné
Thales Canada Inc (Montreal, QC)
Engineering
Information and Communications Technology
Université Laval
Accelerate