Computer Sciences Colloquium - No data? no problem

I will describe two strategies for building robust models when no suitable supervised data is available. The first imodel was built in collaboration with Barak Itkin and Nachum Dershowitz and is in the field of computerized archeology, as part of the ArchAIDE project. The second model generates music and was done by the Facebook AI Research team at Tel Aviv in collaboration with Noam Mor, Adam Polyak, and Yaniv Taigman. 

(i) We are able to learn to identify images of pottery sherds from synthetic 3D data. The method can handle the varying shapes in which a 3D model can break, and manages to classify the potsherds by the shape of the fracture alone. 
(ii) We are able to perform convincing translations across musical instruments, genres, and styles, including musical domains that were not seen during training. The method is completely unsupervised and does not employ matched samples between domains or musical transcriptions.