User Tools
**This is an old revision of the document!** ----
====== Idea ====== MIR techniques rely upon accurate representations of acoustic content in order to produce high-quality results. Over the past few decades, most research has operated on hand-crafted features, which work well up to a point, but may discard important information from the representation, thereby degrading performance. In recent years, deep neural networks have emerged as an effective approach to automatically learn representations of complex signals such as images, video, and speech. Most recently, supervised (discriminative) training of deep networks has been demonstrated to outperform comparable unsupervised methods (such as restricted boltzmann machines). However, in order to do supervised training of features, we require a large pool of accurately labeled data. While this is relatively easy to come by for images, it can be problematic in music. We propose to use artist recognition as a supervised proxy task for training deep representations of musical content. There are two key motivating factors to this idea: - Even if meta-data/tags are unavailable for a particular track, an artist identifier is almost always available; thus it becomes easier to obtain a large-scale training set for discriminative feature learning. - If we build features which can accurately characterize the acoustic signature of an artist, those features may well generalize to other tasks, such as semantic annotation or instrument recognition. ====== Implementation ====== ====== Our stuff ====== * [[https://github.com/bmcfee/deep-artists|Source code]] * [[https://github.com/bmcfee/deep-artists/wiki/Model-architecture|Model architecture]] ====== Resources ====== * [[http://deeplearning.net/tutorial/contents.html|Deep learning tutorial]] * [[https://github.com/Theano/Theano|Theano]] * [[https://github.com/bmcfee/librosa|LibROSA]]
