[Paper Review] Ask, Attend and Answer: Exploring Question-Guided Spatial Attention for Visual Question Answering

Huijuan Xu, Kate Saenko|arXiv (Cornell University)|Nov 17, 2015

Multimodal Machine Learning ApplicationsComputer Science30 references105 citations

TL;DR

This paper proposes the Spatial Memory Network for Visual Question Answering (SMem-VQA), a multi-hop memory network with spatial attention that improves VQA by explicitly modeling spatial inference through question-guided attention over image regions. The model achieves state-of-the-art performance on VQA and DAQUAR datasets, with a 2.35% improvement over the iBOWIMG baseline on the VQA test-standard split, and enables interpretable visualization of attention-based reasoning steps.

ABSTRACT

We address the problem of Visual Question Answering (VQA), which requires joint image and language understanding to answer a question about a given photograph. Recent approaches have applied deep image captioning methods based on convolutional-recurrent networks to this problem, but have failed to model spatial inference. To remedy this, we propose a model we call the Spatial Memory Network and apply it to the VQA task. Memory networks are recurrent neural networks with an explicit attention mechanism that selects certain parts of the information stored in memory. Our Spatial Memory Network stores neuron activations from different spatial regions of the image in its memory, and uses the question to choose relevant regions for computing the answer, a process of which constitutes a single "hop" in the network. We propose a novel spatial attention architecture that aligns words with image patches in the first hop, and obtain improved results by adding a second attention hop which considers the whole question to choose visual evidence based on the results of the first hop. To better understand the inference process learned by the network, we design synthetic questions that specifically require spatial inference and visualize the attention weights. We evaluate our model on two published visual question answering datasets, DAQUAR [1] and VQA [2], and obtain improved results compared to a strong deep baseline model (iBOWIMG) which concatenates image and question features to predict the answer [3].

Motivation & Objective

• To address the lack of explicit spatial reasoning in existing VQA models that rely on global image features and recurrent networks.
• To enable visual question answering models to perform multi-step spatial inference by modeling object locations and relationships through a memory network architecture.
• To design a question-guided spatial attention mechanism that aligns individual words in the question with specific image regions for fine-grained evidence gathering.
• To evaluate the model's reasoning process through synthetic questions requiring spatial inference and visualize attention weights to interpret model behavior.
• To achieve improved performance on standard VQA and DAQUAR benchmarks compared to strong baselines, including iBOWIMG and DPPnet.

Proposed method

• The model uses a memory network that stores feature activations from different spatial regions of the image as memory vectors, enabling spatial attention over visual features.
• In the first hop, word-level attention is applied by computing correlation scores between each word embedding and image patch features, enabling fine-grained alignment between question words and image regions.
• In the second hop, the model uses the entire question embedding and the attended features from the first hop to compute a refined attention map, selecting more precise visual evidence for answer prediction.
• The network is trained end-to-end using a cross-entropy loss on answer prediction, with attention weights learned via backpropagation.
• A third hop is explored but found not to improve performance, indicating diminishing returns beyond two hops.
• The model is evaluated on the VQA and DAQUAR datasets, with attention weights visualized to interpret spatial inference processes.

Experimental results

Research questions

• RQ1 Can a memory network with spatial attention learn to perform multi-hop reasoning over image regions to answer visual questions?
• RQ2 Does question-guided spatial attention improve performance on VQA compared to global image feature models?
• RQ3 Can the model's attention mechanism be visualized to reveal logical inference steps based on spatial relationships?
• RQ4 How does the performance of a two-hop attention mechanism compare to one-hop and three-hop variants?
• RQ5 Can synthetic questions requiring spatial inference effectively probe and validate the reasoning capabilities of the model?

Key findings

• The SMem-VQA Two-Hop model achieves a test-standard accuracy of 58.24% on the VQA dataset, representing a 2.35% improvement over the iBOWIMG baseline (55.89%).
• On the DAQUAR dataset, the SMem-VQA Two-Hop model achieves 79.05% accuracy, outperforming the iBOWIMG baseline (76.55%).
• The model demonstrates superior per-answer category accuracy, particularly on complex spatial reasoning categories, indicating better generalization for spatial questions.
• Visualization of attention weights confirms that the model learns to align specific question words (e.g., 'cat', 'basket') with corresponding image regions, enabling interpretable reasoning.
• The addition of a second hop improves performance over the one-hop version (56.56% on VQA test-standard), showing that multi-hop reasoning enhances spatial inference.
• A third hop does not improve performance, suggesting that two hops are sufficient for effective spatial attention in this setup.