Real-world time series data are rich in sequential and structural patterns. Music, for example, often have a multi-level organization of musical events, with higher-level building blocks made up of smaller recurrent patterns. For computers to understand and process such time series data, we need a mechanism to uncover the underlying structure. Toward this goal, we propose and formulate a number of music puzzle games to test the ability of contemporary neural network models to mine sequential patterns. In essence, these games require a model to correctly sort a few multisecond, nonoverlapping music fragments, either from the same song or not. In the training stage, we learn the model by sampling multiple fragment pairs from the same songs and seeking to predict whether a given pair is consecutive and is in correct chronological order. As no manual labels are needed, it is an unsupervised (more specifically, self-supervised) learning problem. On the basis of state-of-the-art Siamese convolutional network, we propose an improved architecture that learns to embed frame-level similarity scores computed from the input fragment pairs into a common space, where fragment pairs of different types can be more easily distinguished. Our experiments show that the resulting model, dubbed as the similarity embedding network (SEN), performs better than competing models across different games, including music jigsaw puzzle, music sequencing, and music medley.
View on arXiv