A model is said to have been trained end-to-end when it learns a direct mapping from input to output, in an optimization process that handles all parameters "at once". This is in contrast to models which are a composition of multiple stages, and therefore require each stage to be optimized independently. E.g. an image classification model which operates on engineered ("hand-crafted") image features is not end-to-end, since the model has no direct access to the input images.