How Machine Learning helps Pandora find the music of the moment

Finding the music of the moment can often be a challenging problem, even for humans with well-versed musical tastes. These challenges further explode into a myriad of complexities when attempting to construct algorithmic approaches for automatic playlist generation. A variety of factors play a role in influencing a listener’s perception of what music is appropriate on a given seed (e.g. musicological, social, geographical, generational), and these factors vary across different contexts and listeners.

Erik Schmidt, Senior Scientist at Pandora will be presenting at the Machine Intelligence Summit in San Francisco, 23-24 March. Erik will present an overview of recommendations at Pandora, followed by a deep dive into the challenges of recommending content.

Nikita Johnson, Founder of RE•WORK interviewed Erik ahead of the summit to hear more about his work at Pandora, what he feels are the leading factors enabling recent developments in machine learning, what excites him about the industry and which industries will be most affected by machine intelligence (MI). Plus, his predictions for MI in the next 5 years.

Tell us more about your work at Pandora

One of the most fascinating areas of research at Pandora is Machine Listening. In the music domain, we seek to develop systems that are capable of automatically understanding the musicological content of an audio signal. These systems rely heavily on supervised machine learning, and Pandora’s Music Genome Project provides the largest and most detailed corpus in the world for performing this work, spanning over 1.5 million analyzed tracks. As a result of this dataset, we have been able to develop incredibly rich and accurate machine listening representations.

At the core of Pandora’s recommendation system is a massive ensemble that spans over 60 recommendation strategies. Most listeners are likely unaware, but we are constantly adding, removing, and tweaking these strategies to improve the listener experience. Where things get really exciting is when we’re able to take things developed on the research side, like machine-predicted musicological attributes, and incorporate them into recommendation systems. It’s possible to get tens of millions of thumbs back within a short period of time, and there is definitely something both exciting and addictive about working at this scale.

The most exciting large-scale project I’ve led to date is the design of Thumbprint Radio, a hyper-personalized station that has had over 25 million listeners. Thumbprint is a multi-genre experience that sequences all of your thumbs from across all of your stations alongside personalized recommendations. It has spun over 7.6 billion tracks and is currently one of the top stations on Pandora.

What do you feel are the leading factors enabling recent advancements in machine learning?

Currently, there is incredible research momentum in machine learning, spanning both academia and industry. While there have been great recent advances, the truth is that many of the most successful techniques being used today have been around (in some form) for decades. What hasn’t existed until recently is the massive data collections and computing resources that have been amassed in industry.  Industry sized datasets have transformed machine learning challenges by providing billions of examples, a scale which was previously unattainable to the academic community. This has been coupled with the decreasing cost of computing resources, which has made it increasingly more feasible to construct massive computing clusters.

What present or potential future applications of machine intelligence excite you most?

It’s always been my passion for music that has driven my personal interests in machine learning, and so I am therefore most excited about what we can do for music discovery. There’s so much that we can do with platforms like Pandora in terms of surfacing the next big artist and placing them in front of the right listener at the right time. Machine Learning tools that can analyze the audio waveform to match songs or artists musicologically with a listener’s likes are certainly helpful in this problem. Another component is leveraging things like social media in order to identify emerging artists which are gaining a following.

Which industries will be most disrupted by machine intelligence?

The reality is that we’re likely to see literally every industry affected by machine intelligence in the coming years. Machines are already producing a great deal of the goods we consume from electronics to cars. Transportation and agriculture are likely soon to come. How we respond to this and how we adapt our society will likely be one of the great challenges that lies ahead.

What developments can we expect to see in machine intelligence in the next 5 years?

Certainly no one will be struggling to find music discovery tools. I think conversational AI will potentially add a really exciting new level of machine interaction in this area. But in the general landscape of machine learning, we’re witnessing amazing progress happening into complex reasoning tasks such as DeepMind’s work in reinforcement learning. Right now, this is mostly focused around gaming, but there are endless applications for this technology. In terms of more concrete tasks, there has also been incredible progress of many into self driving cars recently, and I expect we will see this become normal.

One of the coolest areas in my opinion is creativity, which is an extremely challenging problem for machines. But we are seeing some exciting efforts in music and art generation which could potentially produce some very cool results in the coming years.

Join experts at the Machine Intelligence Summit and register your place here: https://re-work.co/events/machine-intelligence-summit-san-francisco-2017