Abstract: I will describe a few different ways to apply deep learning techniques to construct generative models for musical language. I will consider some parallels between music generation and language models, and explore the use of different architectures including auto encoders and recurrent neural networks. While most previous work on music generation has effectively focused on creating musical scores, I will also show how we can create piano performances directly by learning a richer language model that includes expressive timing, dynamics and articulation concurrently with the notes.
Bio: After completing his undergraduate degree in math at Dalhousie University in Canada, Sageev Master’s and PhD in Computer Science at the University of Toronto under the supervision of Geoffrey Hinton and Demetri Terzopoulos. Sageev’s research interests include probabilistic generative models and machine learning and deep learning architectures, with an emphasis on creative applications. As a pianist, Sageev has performed with orchestras and at jazz festivals across Canada. Since 2016, he has been a Visiting Research Scientist at Google Brain, and in 2018 he will be joining the Vector Institute and Dalhousie University.
1131 Kemper Hall