Definition
Plain language
Training one system from raw input to final answer in one shot, instead of hand-engineering each step in between.
As stated in the literature
A training paradigm in which a single model is optimized directly on a downstream objective, replacing hand-crafted feature pipelines with learned representations.
Why it matters: It lets the model discover features humans might miss and removes brittle hand-crafted stages that often become the weak link in a pipeline.
For example, instead of writing separate code to detect edges, find shapes, and identify objects in a photo, you train one neural network that takes the raw pixels and outputs the label directly.
Heard on the show
“Both got obliterated by end-to-end learning.”Episode 060 — When Splitting One Model Across Three Agents Doubles Its Accuracy