Transfer learning has been shown to improve training times in other modes (such using an image classification model to initialize an NLP model) over randomly initialized values.