Generally Intelligent

Episode 28: Sergey Levine, UC Berkeley, on the bottlenecks to generalization in reinforcement learning, why simulation is doomed to succeed, and how to pick good research problems

Mar 1, 2023

01:34:49

Creator website

Episode guests

Sergey Levine

Highlights

AI Chapters

Episode notes

Intrinsic Motivation and Reinforcement Learning: Minimizing Surprise

02:30

RL challenges and the importance of architecture selection in large-scale offline RL

02:26

The Role of Labels and Format in Model Learning

01:44

The Benefit in Science of Testing Extreme Designs

01:24

Introduction

2min

Computer Graphics Research

2min

Postdocs - What Happened After That?

2min

Is There a Future for Computer Animation?

2min

I Think Jason Pang's Robotics Demo Is the Work

3min

Using Convolutional Network Techniques to Control Robot Motion

4min

The Challenges of N10 Neural Nets in Robotics

3min

Deep Learning

2min

How to Scale Up Deep Robotic Learning

2min

10.

The Scale of the Robotics Data Set

2min

11.

The Importance of Large Data Sets in Robotics

2min

12.

Robots That Are Like Humans?

2min

13.

Goal Condition DRL for Robotic Control

4min

14.

Is There a Future for Vision Language Models?

2min

15.

Is There a Higher Level Goal?

3min

16.

Picking Research Questions

2min

17.

Machine Learning Systems That Emulate Humans Through Imitation Learning Through Supervised Learning

2min

18.

Offline Reinforcement Learning - What's Next?

2min

19.

Reinforcement Learning - The Basic Problem With Offline RL

2min

20.

Is There a Principle to It?

2min

21.

Offline Deep Learning

4min

22.

The Intuition for Regression Learning (RL)

2min

23.

Large Scale Regression Learning (RL) in Atari Games

2min

24.

How to Improve Language Models Through Value-Based Reinforcement

2min

25.

Is There a Middle Ground to Scaling Up Reaper Learning?

3min

26.

The Importance of Robotics Versus Not?

2min

27.

Is It Sufficient to Just Build Larger Language Models?

3min

28.

Is There an Alternative to Offline RL?

3min

29.

Using Language Models to Build Robotic Robotic Systems

2min

30.

General and Reusable Robotic Control Models

3min

31.

Is Lifelong Robotic Learning Really the New Normal?

2min

32.

What's the Most Influential Research in Robotics?

3min

33.

The Relationship Between Inference and Sequence Models and Control

2min

34.

Metal Learning

3min

35.

Predictive Modeling

2min

36.

The Robotics Experiment at Google

2min

37.

Is There a Difference Between Intrinsic Motivation and Reinforcement Learning?

2min

38.

Is There a Future in Robotic Learning?

4min

39.

Is Language a Prerequisite for Artificial Intelligence?

2min

40.

Robotics Research - Language Models Hit Walls

2min

41.

Generally Intelligent Podcast

2min

Sergey Levine, an assistant professor of EECS at UC Berkeley, is one of the pioneers of modern deep reinforcement learning. His research focuses on developing general-purpose algorithms for autonomous agents to learn how to solve any task. In this episode, we talk about the bottlenecks to generalization in reinforcement learning, why simulation is doomed to succeed, and how to pick good research problems.

Remember Everything You Learn from Podcasts

Save insights instantly, chat with episodes, and build lasting knowledge - all powered by AI.