Using Sentences as Semantic Representations in Large Scale Zero-Shot Learning

Zero-shot learning aims to recognize instances of unseen classes, for which\nno visual instance is available during training, by learning multimodal\nrelations between samples from seen classes and corresponding class semantic\nrepresentations. These class representations usually consist of either\nattributes, which do not scale well to large datasets, or word embeddings,\nwhich lead to poorer performance. A good trade-off could be to employ short\nsentences in natural language as class descriptions. We explore different\nsolutions to use such short descriptions in a ZSL setting and show that while\nsimple methods cannot achieve very good results with sentences alone, a\ncombination of usual word embeddings and sentences can significantly outperform\ncurrent state-of-the-art.\n