chunked/nltk_chunking/model_doc_clip/chunk_7.txt

After pre-training, natural language is used to reference
learned visual concepts (or describe new ones) enabling zero-shot transfer of the model to downstream tasks.