This paper studies the problem of associating images with descriptive sentences by embedding them in a common latent space. We are interested in learning such embeddings from hundreds of thousands or millions of examples. Unfortunately, it is prohibitively expensive to fully annotate this many training images with ground-truth sentences. Instead, we ask whether we can learn better image-sentence embeddings by augmenting small fully annotated training sets with millions of images that have weak and noisy annotations (titles, tags, or descriptions). After investigating several state-of-the-art scalable embedding methods, we introduce a new algorithm called Stacked Auxiliary Embedding that can successfully transfer knowledge from millions of weakly annotated images to improve the accuracy of retrieval-based image description.