Ask, Attend and Answer: Exploring Question-Guided Spatial Attention for Visual Question Answering (1511.05234v2)

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].

Exploring Spatial Attention for Visual Question Answering

The paper "Ask, Attend and Answer: Exploring Question-Guided Spatial Attention for Visual Question Answering" by Huijuan Xu and Kate Saenko investigates the problem of Visual Question Answering (VQA) by introducing a novel model termed the Spatial Memory Network (SMem-VQA). The model leverages a spatial attention mechanism to bridge the gap between image and language understanding, thereby improving the accuracy of VQA systems.

Overview of the SMem-VQA Model

The core contribution of the paper is the Spatial Memory Network, which incorporates explicit spatial attention through a memory network architecture to tackle the VQA task. Traditional deep image captioning methods, such as convolutional-recurrent networks, often fail to adequately model spatial inference. The proposed SMem-VQA addresses this limitation by storing neuron activations corresponding to different spatial regions of an image in its memory. The question guides the model to select relevant regions for computing the answer, performing a process referred to as a "hop."

Key Components

1. Spatial Memory Network Architecture:
   • The model stores convolutional neural network (CNN) activations from different spatial regions.
   • It employs a recurrent neural network (RNN) with an explicit attention mechanism to select relevant parts of the stored information based on the question.
2. Two-Hop Attention Mechanism:
   • The first hop aligns individual words in the question with image patches to guide the attention mechanism.
   • The second hop considers the whole question to refine the visual evidence gathered from the first hop.

Methods and Results

Synthetic Data Experiments

To understand spatial inference by the model, a series of synthetic questions requiring spatial reasoning were designed. The experiments demonstrated that SMem-VQA can learn logical inference rules and accurately select relevant image regions based on the question:

• Absolute Position Recognition: The model accurately identifies an object's fixed location in an image.
• Relative Position Recognition: The model correctly interprets the position of one object relative to another.

Benchmark Dataset Results

The model was evaluated on two standard datasets, DAQUAR and VQA, with the following results:

• DAQUAR Dataset:
  • One-Hop SMem-VQA: Achieved an accuracy of 36.03%.
  • Two-Hop SMem-VQA: Improved accuracy to 40.07%.
  • Both outperforming other deep learning models and a strong baseline (iBOWIMG).
• VQA Dataset:
  • One-Hop SMem-VQA: Achieved competitive results with an overall accuracy of 56.56% on the test-dev set.
  • Two-Hop SMem-VQA: Further boosted accuracy to 57.99% on test-dev and 58.24% on test-standard, outperforming the baseline model and other contemporary models.

Implications and Future Directions

This research highlights the importance of incorporating spatial attention in VQA systems for better interpreting and answering visual questions. By allowing the network to focus on relevant image regions based on the question context, the model becomes more capable of performing intricate reasoning tasks that are vital for VQA.

Theoretically, this work paves the way for more advanced and interpretable AI systems capable of joint image and language understanding. Practically, the proposed two-hop attention mechanism can be extended to other multimodal tasks requiring spatial and temporal reasoning.

Conclusion

The Spatial Memory Network represents a significant advancement in the domain of Visual Question Answering by effectively integrating spatial attention mechanisms. The strong performance across synthetic and standard datasets exemplifies its potential. Future work can explore additional hops or alternative attention mechanisms, leveraging pre-trained embeddings and external data to further enhance VQA models.

By providing a detailed examination of spatial attention within VQA and demonstrating its efficacy, this paper contributes valuable insights and methodologies for advancing AI in the intersection of computer vision and natural language processing.