Human-like attention as a supervisory signal to guide neural attention has shown significant promise but is currently limited to uni-modal integration - even for inherently multimodal tasks such as visual question answering (VQA). We present the Multimodal Human-like Attention Network (MULAN) - the first method for multimodal integration of human-like attention on image and text during training of VQA models. MULAN integrates attention predictions from two state-of-the-art text and image saliency models into neural self-attention layers of a recent transformer-based VQA model. Through evaluations on the challenging VQAv2 dataset, we show that MULAN achieves a new state-of-the-art performance of 73.98% accuracy on test-std and 73.72% on test-dev and, at the same time, has approximately 80% fewer trainable parameters than prior work. Overall, our work underlines the potential of integrating multimodal human-like and neural attention for VQA.
@misc{pub12057,
author = {
Sood, Ekta
and
Kögel, Fabian
and
Müller, Philipp
and
Thomas, Dominike
and
Bâce, Mihai
and
Bulling, Andreas
},
title = {Multimodal Integration of Human-Like Attention in Visual Question Answering},
series = {arXiv},
year = {2021},
volume = {arXiv:2109.13139}
}
German Research Center for Artificial Intelligence Deutsches Forschungszentrum für Künstliche Intelligenz
