Explore the concept of AI-Generating Algorithms (AI-GAs) in this 42-minute conference talk from GECCO 2021. Delve into the three essential pillars for creating powerful, general AI: meta-learning architectures, meta-learning learning algorithms, and generating effective learning environments. Discover why AI-GAs present a unique opportunity for the evolutionary reinforcement learning community, including researchers focused on Quality Diversity and Open-Endedness. Examine successful examples of combining evolutionary and mainstream RL ideas, such as the Go-Explore and POET algorithms. Gain insights into future research directions and understand how this approach could revolutionize the field of artificial intelligence. Learn about the challenges of catastrophic forgetting, the importance of intrinsic motivation, and the potential of meta-learning in continual learning scenarios.
AI-Generating Algorithms - A Unique Opportunity
Association for Computing Machinery (ACM) via YouTube
Overview
Syllabus
Intro
Manual Path to AI
Clear Machine Learning Trend: Hand-designed pipelines are ultimately outperformed by learned solutions
Al-Generating Algorithms
Synthetic Petri Dish A Novel Surrogate Model for Rapid Architecture Search
Meta-learn learning algorithms
Meta-Learning Algorithms
Pillar 2: Meta-Learning
Catastrophic Forgetting
Proposal: Use meta-learning to learn to continually learn
Traditional Neuromodulation
Learned Sparsity
ANML Implications
Pillar 2: Meta-learn learning algorithm
A Paradox
Key for Science & Technological Innovation: Generating Problems, Goal Switching
Qualitatively Different
Intrinsic Motivation
Derailment
Avoids Detachment By Remembering Promising Exploration Stepping Stones
Results on Montezuma's Revenge
Pitfall
Go-Explore: Implications
What's missing?
Traditional ML
Paired Open-Ended Trailblazer (POET)
Task: Obstacle Courses
Direct Path Curriculum Fails
Enhanced POET
POET: Future Work
Overall Conclusions
Taught by
Association for Computing Machinery (ACM)
