Schmidhuber revisits his 2020 talk on AI that learns how to learn
Schmidhuber says many modern meta-learning systems rely on humans to define when trials start and end, while his self-improving systems have learned to redefine those boundaries since 1994.
TLDR
Schmidhuber shared his 2020 talk, “Meta Learning Machines in a Single Lifelong Trial,” arguing that it remains relevant to discussions of recursive self-improvement. He says his systems have learned to redefine trial boundaries since 1994—a choice he says humans still make for many modern meta-learners. The talk’s abstract explores algorithms that learn better learning algorithms and traces his research on that question to a 1987 thesis.
Combined views
1 Source, first seen ago