Many success stories involving deep neural networks are instances ofsupervised learning, where available labels power gradient-based learningmethods. Creating such labels, however, can be expensive and thus there isincreasing interest in weak labels which only provide coarse information, withuncertainty regarding time, location or value. Using such labels often leads toconsiderable challenges for the learning process. Current methods forweak-label training often employ standard supervised approaches thatadditionally reassign or prune labels during the learning process. Theinformation gain, however, is often limited as only the importance of labelswhere the network already yields reasonable results is boosted. We proposetreating weak-label training as an unsupervised problem and use the labels toguide the representation learning to induce structure. To this end, we proposetwo autoencoder extensions: class activity penalties and structured dropout. Wedemonstrate the capabilities of our approach in the context of score-informedsource separation of music.
展开▼
