19-06-20 Forms of Recursive Self Improvement

Category: Idea Lists (Upon Request)

Read the original document

<!-- gdoc-inlined -->

Some are much more dangerous than others. Currently, all of these peter out at sub-general problem solving.

1. Program Induction
2. Self-supervised learning (ex., self-play in deep rl)
3. Learned loss functions (ex., evolved policy gradient or learning unsupervised learning rules)
4. Learned Algorithms (ex., godel machine)
5. Learned Architectures (ex., neural architecture search)
   1. Learned architecture components (ex., group / guage equivariate cnns, learned activation functions)
6. Learned Optimizers
7. Learned Learning Environments / Tasks
8. Godel Machine
9. AIXI

Source: Original Google Doc