Visual Question Answering Optimized Framework using Mixed Precision Training

Thanks to the emergence and continued devel-opment of machine learning, particularly deep learning, the research on visual question and answer, also known as VQA, has advanced dramatically, with great theoretical research significance and practical application value. This field of study makes use of multimodal learning, computer vision, and natural language processing techniques. Except for a few academics who presented different types of optimized bi-linear fusion approaches that integrate text and image characteristics in an efficient way, there haven’t been many efforts to optimize the VQA framework. In order to optimize the VQA problem, we offer a unique Visual Question Answering framework in this research. Because both 16-bit and 32-bit floating points provide automatic mixed precision, deep learning architectures can now be optimized with less computation and execution time. Using the VQA 2.0 and CLEVR datasets, the proposed framework has been tested against two models. In terms of overall accuracy and execution time, the experimental findings demonstrated a significant improvement.