Exploring Uncertainty Measures for Image-caption Embedding-and-retrieval Task

With the significant development of black-box machine learning algorithms, particularly deep neural networks, practical demand for reliability assessment is rapidly increasing. On basis concept that “Bayesian knows what it does not know,” uncertainty network outputs has been investigated as a measure classification and regression tasks. By considering an embedding task task, several existing studies have quantified embedded features improved retrieval performance cutting-edge models by model averaging. However, in image-caption embedding-and-retrieval tasks, well-known samples are always easy to retrieve. This study shows method poor investigates another aspect We propose posterior which can accurately assess results. The consistent two measures observed with different datasets (MS-COCO Flickr30k), deep-learning architectures (dropout batch normalization), similarity functions. To best our knowledge, this first perform on