Generally Intelligent

Episode 25: Nicklas Hansen, UCSD, on long-horizon planning and why algorithms don't drive research progress

Dec 16, 2022

01:49:18

Creator website

Highlights

AI Chapters

Episode notes

Quantifying generalization of RL algorithms on DeepMind Control Suite

04:15

The importance of unsupervised adaptation at test time

03:02

Underrated Techniques and Collaborating Across Fields in RL Research

01:22

The Role of Augmentations in Perception and RL

04:07

Introduction

2min

Do You Have a PhD in Computer Science?

5min

Generalization in Computer Vision

5min

Machine Learning

6min

The Risk of Cataclysmic Rememberting When You Reverse the Environment

4min

RL Research - What Tasks Would You Want to Use?

2min

How Many Samples Do You Need With Rewards?

2min

Using Data Augmentation as an Alternative to Domain Randomization?

4min

Using Data Augmentation to Improve Sample Efficiency

2min

10.

TD Learning - Data Augmentation to Auxiliary Objective

2min

11.

How Do You Determine the Data Augmentation Satisfaction Point?

3min

12.

Is RL Really Hard to Measure?

2min

13.

How to Stabilize High Variance Q Targets With Data Augmentation?

5min

14.

Then, What Happened to the Model Predictive Control?

6min

15.

How Much More All Time Does It Take Kind of First Step Than a Model Free Algorithm?

4min

16.

How Do You Think You Can Predict the Future?

2min

17.

Predicting the Future

2min

18.

Do You Need to Model Everything or Do You Just Model Them?

6min

19.

Is There a Way to Make Humans Flexible?

2min

20.

Are You Using Pre-Trained Policies for Your Research?

4min

21.

Is It Possible to Generate a Neural Network Model?

2min

22.

Are You Thinking About Playing Around With Pre-Trained Encoders in RL?

6min

23.

Simp to Real

3min

24.

Is There a Sensor Problem With Home Robotics?

2min

25.

Is RL a Good Idea for Local Motion Robots?

2min

26.

Is There a Future for RL?

5min

27.

Is Machine Learning a General Algorithm?

3min

28.

What Do You Think Makes a Great Researcher?

5min

29.

Is There a Future for AI?

4min

30.

Is There a Culture That Drives the Technology?

3min

31.

Is There a Lot of Research in ML Models?

6min

Nicklas Hansen is a Ph.D. student at UC San Diego advised by Prof Xiaolong Wang and Prof Hao Su. He is also a student researcher at Meta AI. Nicklas' research interests involve developing machine learning systems, specifically neural agents, that have the ability to learn, generalize, and adapt over their lifetime. In this episode, we talk about long-horizon planning, adapting reinforcement learning policies during deployment, why algorithms don't drive research progress, and much more!

Remember Everything You Learn from Podcasts

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