The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Music & AI Plus a Geometric Perspective on Reinforcement Learning with Pablo Samuel Castro - #339
Jan 16, 2020
Pablo Samuel Castro, a Staff Research Software Developer at Google, shares his journey blending music and reinforcement learning. He discusses the innovative Lyric AI project, which uses multiple models to generate song lyrics that maintain creativity and coherence. The conversation also delves into the geometric perspectives in reinforcement learning, enhancing optimal policy formation, and exciting applications in banking to improve interbank payments. Castro’s insights highlight the importance of human feedback and interdisciplinary approaches in advancing AI.
AI Snips
Chapters
Transcript
Episode notes
Music and Academia
- Pablo Samuel Castro pursued his degrees at McGill while playing in a band.
- He prioritized music over potentially more prestigious academic paths.
Initial Lyric Model Shortcomings
- Pablo Samuel Castro's initial lyric model, based on Andrej Karpathy's work, proved to be uninteresting.
- It overused common phrases and lacked specific nouns, according to artist David Usher.
Multiple Models for Lyrics
- Training a single language model to be coherent, creative, and structurally sound for lyrics is challenging.
- Using multiple models, each focusing on a specific aspect, can be more effective.