Dear Samet: This is a great post. Thanks! One question here. If we treat the mapping from image embedding (4096) to word2vec embedding (300) as a regression problem, rather than a classification problem. Do you think if this idea works?