Skip to Content

Luke Barrington (CSE Colloquium)

''Machines that Understand Music"

Luke Barrington
UCSD

Monday, January 9, 2012, 11:00 am
EBU3B, Room 1202

Abstract

Computer audition is analogous to computer vision in that we try to teach machines to hear. One particular audio domain that is both challenging and exciting is the understanding of musical signals. In this talk, I will discuss machine learning models of music that endeavor to analyze audio waveforms and extract meaningful patterns. The models learn how these patterns correspond to semantic descriptions, like "funky party tunes", "aggressive metal with shredding guitar solos" or "mellow electro that's good to listen to while writing code". Such a mapping between words and music powers a range of applications, including music search engines, music recommendation algorithms, song segmentation and audio visualization.

The second half of the talk will consider the problem of how to collect a set of reliably-labeled music, suitable for training our computer audition models. In particular, I will describe Herd It, a musical game on Facebook, that has collected hundreds of thousands of human descriptions of music. By combining crowdsourced data collection with active learning machine models, "game-powered machine learning" achieves scalability and reliability sufficient to label every song on the web.

Short biography

Luke Barrington's Ph.D thesis on "Machines that Understand Music" describes a new approach to extracting information from massive datasets by combining crowdsourced human intelligence with automatic machine learning. As a founder of the UCSD Computer Audition Laboratory and of Music Search, Inc., he applied this work to analyze and recommend millions of songs on the web. As a member of the Valley of the Khans project, he took the science of crowdsourcing into a new arena by engaging thousands of virtual explorers from around the world to guide him as he traveled to Mongolia on a National Geographic expedition to search for the tomb of Genghis Khan. Luke currently serves as the founder and CTO of Tomnod, inc. where he and his team continue to advance both crowdsourced and algorithmic methods to extract rapid, relevant, reliable information from geospatial imagery. Applications of Tomnod's technology include government surveillance, disaster damage mapping and rapid response to humanitarian crises. When not mining the web or saving the world, Luke is an avid musician and wails on the guitar.



by Dr. Radut.