Uncertainty-based loss function weighting for multi-task learning