Vector space modeling on music data

04:30 PM - 05:25 PM on July 16, 2016, Room CR4

Tim Schmeier

Audience level:

intermediate

Watch:

https://www.youtube.com/watch?v=jjO1gOH-BW4

Description

Vector space modeling is a popular machine learning approach with a wide variety of applications, including product recommendations, information retrieval, and image classification. In this talk we will go into detail of some of our innovative approaches to building these vector spaces on music data at iHeartRadio. We will also give several real-world examples of how we use these models, highlighting their power, flexibility, and simplicity.

Abstract

Vector space modeling is a popular machine learning approach with a wide variety of applications, including product recommendations, information retrieval, and image classification. These models are common because they map large complex data spaces into compact simple ones, allowing intelligent decisions to be made very quickly.

In this talk we will go into detail of how we build these models on music data using large-scale machine learning in Spark and deep neural networks in Theano. Using real-world examples, we will show how a small number of simple operations can enable extremely powerful and flexible functionality.