Generalized Hadamard-Product Fusion Operators for Visual Question Answering

TL;DR

This paper introduces a generalized class of multimodal fusion operators for visual question answering, demonstrating that specific instantiations improve accuracy and suggesting potential for discovering even better operators through architecture search.

Contribution

The paper proposes a generalized framework for multimodal fusion operators in VQA, including novel components like Nonlinearity Ensembling and Feature Gating, achieving improved accuracy.

Findings

Achieved 1.1% improvement in VQA 2.0 test-dev accuracy.

Identified that specific instantiations outperform baseline fusion methods.

Proposed a generalized fusion operator class as a search space for future architecture optimization.

Abstract

We propose a generalized class of multimodal fusion operators for the task of visual question answering (VQA). We identify generalizations of existing multimodal fusion operators based on the Hadamard product, and show that specific non-trivial instantiations of this generalized fusion operator exhibit superior performance in terms of OpenEnded accuracy on the VQA task. In particular, we introduce Nonlinearity Ensembling, Feature Gating, and post-fusion neural network layers as fusion operator components, culminating in an absolute percentage point improvement of $1.1\%$ on the VQA 2.0 test-dev set over baseline fusion operators, which use the same features as input. We use our findings as evidence that our generalized class of fusion operators could lead to the discovery of even superior task-specific operators when used as a search space in an architecture search over fusion operators.