Tree-loss function for training neural networks on weakly-labelled datasets | IEEE Conference Publication | IEEE Xplore