They have applied it separately to speech, text and images where it outperformed the previous best single-purpose algorithms for computer vision and speech.