Deep Learning for Music Information Retrieval I: How do Neural Networks Learn Music?
Workshop Date:
Sun, 07/01/2018 - Fri, 07/06/2018
Tuition scholarships available for applicants commited to advancing diversity in STEM. Fill out this form by May 1st.
Instructors: Irán Román, Kitty Shi
An exploration of the mathematical principles that make Neural Networks learn from data. Students will build different Neural Network architectures in Julia (from scratch) and in Python with tensorflow*. We will cover Feedforward, Recurrent and Convolutional Models. Models will be trained to solve Timbre detection, Genre classification, and Natural Language Processing tasks. Students develop their own original research project using Deep Learning. This course is meant for individuals who want to understand how neural networks work. Prerequisites: (1) multivariate calculus and (2) programming proficiency.
*Students may use other toolboxes if they are already familiar with them.
Syllabus:
Day 1: Audio Features and Classification Algorithms
Day 2: Decomposing and building a Softmax Neural Network from scratch, and training it with music.
Day 3: Obtaining Spectral Features from raw audio using Convolutional Neural Networks.
Day 4: Natural Language Processing in Music using Recurrent Neural Networks.
Day 5: Make it your own: Proposing and developing a Final Project with Neural Networks.
Prerequisites:
- CCRMA MIR workshop (any year) or consent from both instructors (iran[at]ccrma_DOT_stanford_DOT_edu & kittyshi[at]ccrma_DOT_stanford_DOT_edu).
- Multivariate calculus and programming experience with Python
About the instructors:
Irán Román studies Computer-based Music Theory at CCRMA and Neuroscience at the Stanford Neuroscience Institute. His dissertation models brain computations during sound perception using artificial neural networks. His industry experience includes the development of artificial neural networks for various applications.
Kitty Shi is a PhD Candidate at CCRMA, and a PhD minor in computer science. She is interested in modeling expressive musical performances. She's been working at Shazam, Adobe for various applications of MIR.
IMPORTANT: Contact the instructor before registering to confirm your eligibility. Attach a copy of your registration or diploma for the CCRMA Music Information Retrieval workshop. Describe your experience with python programming (preferably include a link to your github page), and college-level math classes at the level of Calculus I or above.
