End-to-end learning
Training a single learned system to map raw sensory input directly to control outputs, rather than splitting perception, planning and control into hand-engineered modules
As defined in World's Top Researcher on AI, LLMs, and Robot Intelligence
Other definitions
An approach where a single neural network maps raw sensor input directly to driving actions, without separate modules for perception, planning, and control.
Lex Fridman Podcast #14 - Kyle Vogt
Where it comes up · 2
- World's Top Researcher on AI, LLMs, and Robot Intelligence2 claims
- Lex Fridman Podcast #14 - Kyle Vogt1 claim
Most important claims · 3 of 3Search all claims for “End-to-end learning” →
- End-to-end learned driving is decades old: an Alvinn-style system in the mid-to-late 1980s (Alvin, circa 1986 or 1987) drove highways using a tiny neural network that took camera input.World's Top Researcher on AI, LLMs, and Robot Intelligence
- Robotics has been historically resistant to learning because physical systems seemed to require explicit programming of physics, and the field took a long time to accept that learning systems can figure out physics themselves; learning is now broadly accepted, but end-to-end learning and the bitter lesson remain contested.World's Top Researcher on AI, LLMs, and Robot Intelligence
- Lex argues the deep learning community's extreme optimism about end-to-end learning, where systems perceive and act directly like humans, should not be criticized because eventually it may be right, even though it contrasts with modular, rigorous approaches.Lex Fridman Podcast #14 - Kyle Vogt