Differentially Private Representation Learning via Image Captioning