Theoretical Goals in Reinforcement Learning

Csaba Szepesvári is a prominent figure in machine learning and artificial intelligence, particularly known for his contributions to reinforcement learning. He is a professor in the Department of Computing Science at the University of Alberta and a Senior Staff Research Scientist at Google DeepMind.

Our goal in life: keep eating Yann Lecun's cherry (inside joke, ask GPT).

This was a wide-reaching conversation with my good friend Csaba. The first 50 minutes is about AI, we give you insight into what AI researchers and practitioners are dealing behind the scenes, a bit of history and practice of AI.

At around 49 minutes we discuss Yann Lecun's view on model-based planning and reinforcement learning, which is one of the most interesting and far-reaching discussion within AI (should we model the world or just learn to reach goals?).

This naturally leads into relevance realization: what to learn predict and what to ignore, which is the most important question of AGI.

The last 20 minutes it turns more personal, we talk about our life as scientists, the metaphysical no man's land (nonreductionist naturalist metaphysics), and my wrestling with Christianity.

If you like this, check out my conversations with Yogi Jaeger https://www.youtube.com/watch?v=5UJ4y2L2qpk, Anna Riedl https://www.youtube.com/watch?v=w2ZiSWZNQsg, and Giuseppe Paolo https://www.youtube.com/watch?v=R7tEd65e2i8.

00:00:00 Intro.

00:02:39 Google Deep Mind and leading the Foundations team. Theory tells what is possible, and how to do it? What does theory add to practice in machine learning?

00:08:58 Probabilistic guarantees. Sorting vs machine learning. Random algorithms and random data.

00:15:57 Theory and practice. Support vector machines, boosting, and neural nets. History of AI. Practice-driven ML.

00:24:34 Neural nets: history and practice. The fiddliness and the robustification. Overfitting. Overparameterization vs classical statistics. Label noise and regularization.

00:35:06 Reinforcement learning. Learning to behave. Learning to collect data.

00:39:10 Why did we get into AI: to understand intelligence and ourselves.

00:41:48 Reinforcement learning: model of an intelligent agent. Theory: wrt a fully informed agent, how much do you lose by having to learn?

00:49:17 To model or not to model? Model-based reinforcement learning.

00:58:18 Latent representation of importance, relevance realization. Steelmanning and criticizing Yann Lecun's cherry metaphor. Relevance vs simplicity. The shiny object syndrome.

01:12:40 The Jaeger - Riedl - Djedovic - Vervaeke - Walsh paper on naturalizing relevance realization.

01:24:12 Fear and technology. Regulation, freedoms, open source.

01:33:00 Framing precedes observation.

01:35:60 Subjectivity of science. Can we be part of the world we study?

01:43:55 The metaphysical no man's land: nonreductionist naturalism.

01:48:00 Why am I not yet a Christian? Omnipotence and lack of embodiment, rituals that focus on our responsibility of what is below.

I, scientist blog: https://balazskegl.substack.com

Twitter: https://twitter.com/balazskegl

Artwork: DALL-E

Music: Bea Palya https://www.youtube.com/channel/UCBDp3qcFZdU1yoWIRpMSaZw

Hosted on Acast. See acast.com/privacy for more information.