Loading presentation...

Present Remotely

Send the link below via email or IM


Present to your audience

Start remote presentation

  • Invited audience members will follow you as you navigate and present
  • People invited to a presentation do not need a Prezi account
  • This link expires 10 minutes after you close the presentation
  • A maximum of 30 users can follow your presentation
  • Learn more about this feature in our knowledge base article

Do you really want to delete this prezi?

Neither you, nor the coeditors you shared it with will be able to recover it again.


Make your likes visible on Facebook?

Connect your Facebook account to Prezi and let your likes appear on your timeline.
You can change this under Settings & Account at any time.

No, thanks

Machine Learning, Music, and You

An exploratory presentation on the polar worlds of Computer Science and Music and how breakthroughs in technology can lead to novel improvements in Music Education.

Bryan Watson

on 28 March 2013

Comments (0)

Please log in to add your comment.

Report abuse

Transcript of Machine Learning, Music, and You

So What is Machine Learning? "The field of study that gives computers the ability to learn without being explicitly programmed."
-Arthur Samuel It's a branch of Artificial Intelligence Not this.... But this.... Sentient (or self-aware) machines/computers as portrayed in movies. Neural Networks Deep Blue Defeated World Chess Champion
Garry Kasparov on May 11, 1997. IBM Watson Made famous by beating former Jeopardy Champions Brad Rutter and Ken Jennings in 2011, IBM Watson is now used to diagnose people for lung cancer. Xbox Kinect Won the 2011 MacRobert Award in engineering for its work with human motion capture. Applications Detecting Credit Card Fraud Improving Search Engines Medical Diagnosis Natural Language Processing Computer Vision IBM Watson - Lung Cancer Detection University of Wisconsin Hospitals - Breast Cancer Detection Voice/Speech Recognition Handwriting Recognition Image/Object Detection "Netflix Prize" http://www.netflixprize.com/ Biological Neural Networks Artificial Neural Networks Machine Learning focuses on designing programs that can adapt and discover new features on their own. So What Can Machine Learning Do? Geoffrey Hinton of Toronto University
Coursera.org - Neural Networks for Machine Learning Neurons are cells in the brain that transmit information through chemical and electrical signals. Each Neuron has a body called the soma (2), a nucleus (3), a tail called the axon (4), and branches called dendrites (1). When many of these neurons (there are around 86 billion in the human brain!) are connected, the brain gains the ability to remember information and make decisions. How all of these neurons are connected determines how we think and act, but scientists have only scratched the surface and many of the brain's inner functions are still unknown. http://www.guardian.co.uk/science/blog/2012/feb/28/how-many-neurons-human-brain The Netflix Prize was a contest in 2007 to design a better collaborative filtering algorithm to predict user ratings for films. The winning team managed to increase the accuracy by 10.6% using a Machine Learning algorithm called a Neural Network. Artificial neurons are modeled in a similar way to their biological cousins. An artificial neuron has a body with an output and some number of inputs. Each input has a weight that is multiplied with the input signal. When the sum of all of these inputs is greater than the neuron's threshold it sends out a signal. The learning happens when you adjust the weights in the input lines, which in turn changes when the neuron "fires." A network of these neurons can thus learn to make a choice. Given an input... ...the network will chose and output Bryan Watson - Senior Presentation Machine Learning, Music, and You Computer programs that behave Intelligently. The problem is that the term Artificial Intelligence has lost credibility. This is partly because of over-promising by scientists and its portrayal in media. The truth is:
Designing intelligent computers is much harder than people originally thought. So what does this have to do with music? Music Composition Music Selection Three Key Areas in Music Music Recognition Shazam Midomi VS. They both "listen" to a segment of music and determine what song it is. compares User records music with program Database compares User plays a segment of music Database of over 8,000,000 songs "Fingerprints" of Songs Fingerprint of segment Each "fingerprint" is made by creating a spectrogram of the song. The trick is to cut out most of the details and only focus on the peaks. compares User sings or hums a melody segment 100% user generated database of songs User recordings of songs Becomes better with each additional song added. How it Works Uses a special search engine called MARS (Multimodal Adaptive Recognition System), which is the first program to match human voice with human voice. The Music Genome Project A ten year project to develop the world's most thorough and sophisticated database of music ever created. Pandora does not, however, use any Machine Learning. Instead, it relies entirely on human input. This method is time consuming and vulnerable to human bias and error. Automatic Musical Pattern Feature Extraction Using Convolutional Neural Network Students at the University of Hong Kong developed program that could look at a song and determine what genre the song was from. Listen to this recording. David Cope Once described Mozart as a composer who "...was able to digest [music] and store it in his database. He could recombine it with other things so that the output would be hardly recognizable." Having trouble finishing his opera, Cope decided to write a program to help him come up with ideas. Database Songs encoded by hand Recombined patterns The program then learns and analyzes the songs for patterns. Experiments in Musical Intelligence (Emmy) Who do you think may have composed this? Emily Howell (1981 - present) Computer program written by David Cope Emmy could soon copy the styles of classical composers. Shazam Ignores: key, tempo, language, and singing quality. Focuses on: pitch variation, rhythm, phonetic content, and speech content. Midomi Database List of musical attributes for each song These 450 attributes are determined by a team of experts 20 to 30 minutes are dedicated to each 4 minutes of song. Song or artist seed Songs with similar attributes First, they converted each song into an image. Then they used a Convolutional Neural Network which specializes in image recognition to learn features like musical texture and rhythm. The result was a program that was over 87% accurate within the general library but performed badly outside of the training sets. Despite the setback, the program has the potential to enhance or supplant Pandora Radio's current methods. Author, composer, scientist The code for each note included
When the note appears
Which voice it belongs to In 1987, Cope finished his opera with Emmy's help. As you can imagine, there was and is debate about whether Emmy's work constitutes as art. He continued to expand on the program's abilities, adding music from many genres, not just classical and baroque. By 1992, Emily Howell (the program's new name) had written 1500 Symphonies, 1000 piano sonatas, and 1000 string quartes. Further Reading Neural Networks http://yann.lecun.com Music http://www.technologyreview.com/view/419223/using-neural-networks-to-classify-music/ Under Construction
Full transcript