IMCN: Improved modular co-attention networks for visual question answering

Many existing Visual Question Answering (VQA) methods use traditional attention mechanisms to focus on each region of the input image and each word of the input question and achieve well performance. However, the most obvious limitation of traditional attention mechanisms is that the module always generates a weighted average based on a specific query. When all regions and words are unsatisfied with the query, the generated vectors, which are noisy information, may lead to incorrect predictions. In this paper, we propose an Improved Modular Co-attention Network (IMCN) by incorporating the Attention on Attention (AoA) module into the self-attention module and the co-attention module to solve this problem. AoA adds another attention process by using element-wise multiplication on the information vector and the attention gate, which are both generated from the attention result and the current context. With AoA, the attended information obtained by the model is more useful. We also introduce an Improved Multimodal Fusion Network (IMFN), which leverages various branches to achieve hierarchical fusion, to fuse visual features and textual features for further improvements. We conduct extensive experiments on the VQA-v2 dataset to verify the effectiveness of the proposed modules and experimental results demonstrate our model outperforms the existing methods.