end-to-end learning

Training a single model, usually a deep neural networkneural network, n.: A kind of computer program that learns a task from examples by tuning thousands to billions of connections between simple number-crunching units, an arrangement loosely inspired by nerve cells in the brain., to map raw inputs directly to final outputs, with all components optimized jointly for the final objective by backpropagationbackpropagation, n.: The method neural networks use to learn from mistakes: it works backward from the error to figure out how much each internal connection contributed, so every connection can be adjusted in the right direction., instead of building a pipeline of separately designed and trained stages. Examples include speech recognizers that map audio straight to text without separate acoustic, pronunciation and language models; neural machine translationmachine translation, n.: Software that renders writing or speech in one language into another language, as online translation services and phone translation apps do. replacing phrase-based systems; and driving or robot policies that map camera pixels to controls. The advantages are that no hand-designed intermediate representation limits the system and that errors do not compound across separately tuned stages. The costs are a large appetite for data, failures that are hard to diagnose because there are no inspectable intermediate outputs, and difficulty injecting domain knowledge or guaranteeing safety, which is why many deployed systems remain hybrids.

We dropped the separate parser and trained the whole system end to end, and accuracy rose six points.