2017/04/02 by Tanmay Gupta, Gupta, Tanmay, Kevin J. Shih +5
Computer Science · #Artificial Intelligence (cs.AI) #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Multimodal Machine Learning Applications #Neural and Evolutionary Computing (cs.NE) #Topic Modeling
paper · pdf · doi:10.48550/arxiv.1704.00260
openalex publication_date 2017/04/02 · openalex created_date 2022/10/01 · openalex updated_date 2026/07/28
An important goal of computer vision is to build systems that learn visual\nrepresentations over time that can be applied to many tasks. In this paper, we\ninvestigate a vision-language embedding as a core representation and show that\nit leads to better cross-task transfer than standard multi-task learning. In\nparticular, the task of visual recognition is aligned to the task of visual\nquestion answering by forcing each to use the same word-region embeddings. We\nshow this leads to greater inductive transfer from recognition to VQA than\nstandard multitask learning. Visual recognition also improves, especially for\ncategories that have relatively few recognition training labels but appear\noften in the VQA setting. Thus, our paper takes a small step towards creating\nmore general vision systems by showing the benefit of interpretable, flexible,\nand trainable core representations.\n