Computer Vision News - March 2017

34 Computer Vision News RE•WORK Summits in S. Francisco Events discussed much higher-level problems, such as a developing an intelligent adversary for computer gaming, and machine learning has focused on more concrete data-driven tasks such as statistical pattern recognition. However, in recent years there has been an explosion of progress in machine learning, which arguably has led to some of the most significant advances on the path to AI thus far. So in many ways the conference theme is very timely. Erik, how does it directly affect your work at Pandora? Erik: At a high-level, Pandora is an artificial intelligence system that presents itself to the listener as a simple agent, delivering a never ending stream of personalized music recommendations. At this level, the goal of the system is to seamlessly operate utilizing limited feedback signals, thus making the product appear almost magical. However, deep below the surface of Pandora’s simple interface is a massive music recommendation engine, spanning over 60 recommenders which utilize numerous machine learning techniques. And what advancements in machine learning will have the biggest impact on the music sector? Erik: Machine learning offers incredibly powerful tools to drive the music discovery process. This means introducing a listener to new music as well as introducing a new artist to the world. Through machine listening we have models that can analyze the musicological characteristics of any audio file in a fraction of a second. With collaborative filtering we harness tools capable of analyzing the behavior of millions of listeners in order to understand personal preferences. What could get very exciting from an artist’s perspective is the ability to match them with an audience even if they are entirely new. Any song placed online could have a musicological analysis instantly. Through identifying taste leaders we could predict an artist’s future fan base by analyzing their early listeners. It’s also possible that intelligent systems will become capable of creating music, which could be quite fascinating as well. However, it’s thought that music has evolved explicitly for the expression of emotions between humans, and as such it seems unlikely that machines will ever take over this area completely. What algorithms provide us is a powerful set of tools to create human experiences that otherwise might not happen. Machine learning in music “It’s also possible that intelligent systems will become capable of creating music”