Moving beyond feature design: Deep architectures and automatic feature learning in music informatics

The short history of content-based music informatics research is dominated by hand-crafted feature design, and our community has grown admittedly complacent with a few de facto standards. Despite commendable progress in many areas, it is increasingly apparent that our efforts are yielding diminishing returns. This deceleration is largely due to the tandem of heuristic feature design and shallow processing architectures. We systematically discard hopefully irrelevant information while simultaneously calling upon creativity, intuition, or sheer luck to craft useful representations, gradually evolving complex, carefully tuned systems to address specific tasks. While other disciplines have seen the benefits of deep learning, it has only recently started to be explored in our field. By reviewing deep architectures and feature learning, we hope to raise awareness in our community about alternative approaches to solving MIR challenges, new and old alike.