Speaker: Hayoung Lyou & Keunwoo Choi
Title: Jazz Composition and Improvisation in Juxtaposition with Large Language Models Time: Apr 15th, 2025, 18:00 - 20:00
Place: N4, 1F #1101 (인문사회과학부동 1층, 시청각실)

Abstract:
This interdisciplinary presentation explores the fascinating parallels between jazz improvisation and large language models (LLMs), examining the creative processes that underlie both musical and computational generation. In the first part, "Variation and Repetition: The Making of 'The Myth of Katabasis (2024)'," jazz composer and pianist Hayoung Lyou will perform the solo piano pieces "Descent Suites (i, ii, iii)" from her recent album. She will discuss her album-making process and, through an interview with Jaekwon Lim, demonstrate how she develops small musical ideas into a complete album using various compositional techniques. The presentation will offer insights into the creative journey from initial concept to finished work, illuminating the artistic decisions that shape jazz composition. The second part, "How Music and Text are Written by Humans and Large Language Models," led by Senior Principal ML Scientist Keunwoo Choi, will examine parallels between musical creation and text generation through LLMs. The presentation will explore reasoning in LLMs, comparing intuitive processes with more deliberate thought. As both musician and AI researcher, Dr. Choi offers a unique perspective that bridges compositional techniques with AI systems. By drawing connections between musical practice and LLM training, and between improvisation and model inference, this talk presents a fresh framework for understanding creativity across human and artificial domains.

Bio: Hayoung Lyou (https://www.hayounglyou.com/)
Hayoung Lyou is a jazz composer and pianist whose recent trio album "The Myth of Katabasis" (2024) was released on Endectomorph Music, featuring Thomas Morgan and Steven Crammer. She previously released her debut album "Metamorphosis" (2020), a quintet project featuring all original compositions. Lyou holds a Master of Music in Jazz Studies from New England Conservatory and a Bachelor of Music in Performance from Berklee College of Music. Her accolades include being a semi-finalist in the International Songwriting Competition and receiving scholarships from New England Conservatory and Berklee. She has performed at notable venues including The Jazz Gallery and Bar Bayeux in New York, and recently conducted a workshop on composition and improvisation for jazz students at the University of Colorado Boulder. As a pianist, she has contributed to numerous recording projects, including Elijah Shiffer's "DADA BEBOP" and Jacob Shulman's albums "High Firmament" and "Ferment Below." 

Keunwoo Choi (https://keunwoochoi.github.io/) 

Dr. Keunwoo Choi is a Senior Principal ML Scientist and Team Lead at Prescient-LM, Genentech, where he leads research and development for internal large language model initiatives while managing a team of ten members. His previous experience includes roles at Gaudio Lab, ByteDance/TikTok, and Spotify, where he focused on music and audio analysis, generation, and separation. Dr. Choi earned his PhD in Electrical Engineering from the Centre for Digital Music at Queen Mary University of London, specializing in music and audio AI. He served as an Adjunct Professor at New York University, teaching "Deep Learning for Media," and will be the General Chair for ISMIR 2025. He is also the creator of "LLMs heart MIR," a tutorial on large language models for music information retrieval. His unique background bridges the domains of music creation and machine learning, informing his perspective on both fields.