Text Descriptions are Compressive and Invariant Representations for Visual Learning

Zhili Feng, Anna Bair, J. Zico Kolter

Modern image classification is based upon directly predicting classes via large discriminative networks, which do not directly contain information about the intuitive visual featur…