Generally Intelligent Nicklas Hansen, UCSD: Long-horizon planning and why algorithms don't drive research progress
18 snips
Dec 16, 2022 AI Snips
Chapters
Transcript
Episode notes
Stumbling into Research
- Nicklas Hansen stumbled into research during college after initially not knowing it was a paid opportunity.
- He was drawn to research by the prospect of contributing impactful work beyond coursework and hobbies.
Generalization Issues
- Computer vision models trained on ImageNet struggled with generalization to similar datasets, a problem also faced by RL models.
- RL research at the time focused on single-task mastery with minimal randomization.
DeepMind Control Experiment
- Nicklas Hansen and his team tested RL algorithms on a new DeepMind Control Suite test set with minor variations.
- The algorithms performed poorly, highlighting overfitting to the original environment due to limited visual experience.