Our new X account is live! Follow @wizwand_team for updates
WorkDL logo mark

KAT: A Knowledge Augmented Transformer for Vision-and-Language

About

The primary focus of recent work with largescale transformers has been on optimizing the amount of information packed into the model's parameters. In this work, we ask a different question: Can multimodal transformers leverage explicit knowledge in their reasoning? Existing, primarily unimodal, methods have explored approaches under the paradigm of knowledge retrieval followed by answer prediction, but leave open questions about the quality and relevance of the retrieved knowledge used, and how the reasoning processes over implicit and explicit knowledge should be integrated. To address these challenges, we propose a novel model - Knowledge Augmented Transformer (KAT) - which achieves a strong state-of-the-art result (+6 points absolute) on the open-domain multimodal task of OK-VQA. Our approach integrates implicit and explicit knowledge in an end to end encoder-decoder architecture, while still jointly reasoning over both knowledge sources during answer generation. An additional benefit of explicit knowledge integration is seen in improved interpretability of model predictions in our analysis.

Liangke Gui, Borui Wang, Qiuyuan Huang, Alex Hauptmann, Yonatan Bisk, Jianfeng Gao• 2021

Related benchmarks

TaskDatasetResultRank
Visual Question AnsweringOK-VQA (test)
Accuracy54.41
296
Visual Question AnsweringOK-VQA
Accuracy50.6
224
Visual Question AnsweringA-OKVQA (test)
Accuracy54.4
79
Visual Question AnsweringOK-VQA v1.0 (test)
Accuracy54.41
77
Visual Question AnsweringEnc-VQA (test)
Single-Hop Accuracy40.9
69
Visual Question AnsweringOK-VQA (val)
Accuracy54.4
47
Visual Question AnsweringInfoSeek (Full)
Accuracy22.3
35
Visual Question AnsweringEncyclopedic-VQA Full
Accuracy33.9
35
Visual Question AnsweringOK-VQA v1.1 (test)
VQA Score54.41
28
Knowledge-based Visual Question AnsweringOK-VQA v1.0 (test)
Accuracy54.41
15
Showing 10 of 12 rows

Other info

Follow for update