Faithful Multimodal Explanation for Visual Question Answering

Motivation

• The paper targets VQA systems whose deep architectures are difficult to comprehend and therefore difficult to trust.Existing visual explanations identify influential regions, while textual explanations provide reasons but do not necessarily show how image regions support the answer.Source: §1 · p. 1
• Multimodal explanations combine relevant image regions with textual explanations whose corresponding words refer to those regions.Figure 1 illustrates this word-to-region correspondence.Source: §1 · p. 1, Figure 1 · p. 1

Approach

• The approach trains an explanation generator on VQA-attended features and filters explanations whose visual grounding is inconsistent with the VQA reasoning process.It uses GradCAM to trace explanations to relevant objects and aligns the explanation and VQA visual explanations during training.Source: §1 · p. 2

Evaluation

• The system evaluates explanation quality with automated and human metrics, including similarity to LIME-influential segments and consistency with gradient-based VQA explanations.These metrics are designed to assess whether explanations reflect influential parts of the underlying decision process.Source: §1 · p. 2

2.1 VQA

• VQA models combine CNN image features and RNN question features to predict answers, often using attention to emphasize important visual information.The paper adopts Bottom-Up-Top-Down attention but replaces object detection with segmentation for more precise boundaries.Source: §2.1 · p. 2
• The paper replaces object detection with segmentation in its adopted attention mechanism to reduce emphasis on irrelevant image portions and obtain more precise object boundaries.The attention operates over high-level objects rather than convolutional features.Source: §2.1 · p. 2

2.2 Visual Explanation

• Visual explanation methods such as GradCAM and attention maps highlight regions affecting a decision but do not explain relationships among those regions.The proposed approach links segmented objects to words in an accompanying textual explanation and filters noisy attention.Source: §2.2 · p. 2

2.3 Textual and Multimodal Explanation

• Textual and multimodal explanation methods add linguistic reasons or connect text with visual regions, but the paper emphasizes referencing objects that actually influenced the system’s decision.This faithfulness criterion distinguishes the paper’s goal from explanations that are merely grounded in the image.Source: §2.3 · p. 2

3 Approach

• The system segments images, predicts answers with a pretrained VQA module, and conditions textual explanations on the question, answer, and VQA-attended features.The VQA parameters remain fixed while the explanation module is trained.Source: §3 · p. 3, §3.2 · p. 3

3.2 VQA Module

• Instance segmentation supplies precise object-level visual features, while the VQA module attends over these objects to produce answer logits.The model retains at most the top V < 80 segmented objects and represents each with a 2048-dimensional vector.Source: §3.2 · p. 3

3.3 Question and Answer Embedding for Explanation Generation

• Question and sampled-answer embeddings are combined with VQA-attended features so the explanation representation directly reflects the VQA process’s visual focus.The resulting joint representation is derived directly from VQA-attended features.Source: §3.3 · p. 4

3.4 Explanation Generation

• The two-layer LSTM explanation module uses visual attention and a source identifier to decide whether each generated word should rely on attended image content or linguistic state.The source identifier balances attention to visual features and the recurrent hidden representation, with supervision based on GloVe similarity above 0.6.Source: §3.4 · p. 4
• Faithful supervision filters human explanations by Grad-CAM cosine similarity with the predicted answer and jointly optimizes explanation, source-identification, and gradient-alignment losses.Explanations with faithfulness below ξ max(0.02 i_t, 1) are filtered, wrong-answer explanations are discarded, and local faithfulness uses L_f = 1 − S_f.Source: §3.4 · p. 5

3.5 Training

• Training uses 29,459 VQA-X question-answer pairs, a frozen VQA module, and faithfulness threshold ξ = 0.3 before optimizing the explanation module.The VQA module is pretrained for 15 epochs, and the explanation module is trained for 25 epochs with Adam.Source: §3.5 · p. 5

3.6 Multimodal Explanation Generation

• The final multimodal explanation links common nouns to segmented objects when the source identifier exceeds 0.5 and the corresponding attention weight exceeds 0.2.Each linked noun is paired with the segmented object receiving the highest attention at the word’s generation step.Source: §3.6 · p. 6

4.1 Textual Explanation Evaluation

• The approach outperforms PJ-X on all automated textual-explanation metrics despite using about half the explanation training data.The comparison uses BLEU-4, METEOR, ROUGE-L, CIDEr, and SPICE, with ablations included.Source: §4.1 · p. 6

4.2 Multimodal Explanation Evaluation

• The approach matches human attention maps more closely than PJ-X by using detailed segmentation and linking visual objects to words in the textual explanation.These design choices reduce background and unnecessary-object attention compared with purely visual explanations.Source: §4.2 · pp. 6–7
• About 70% of human evaluations are positive and about 45% are strongly positive for both answer support and word-to-region link quality.The evaluation used 1,000 correctly answered questions and Likert-scale judgments of support and linking quality.Source: §4.2 · p. 7

4.3 Faithfulness Evaluation

• The model links an average of 1.6 explanation words to image segments, indicating that generated textual explanations are grounded in objects detected by the VQA system.This faithfulness analysis uses explanations from 1,000 correctly answered test questions.Source: §4.3 · p. 7
• LIME’s influential objects match explanation-linked objects with an average cosine similarity around 0.7, while faithful supervision and local-faithfulness enforcement improve the metric over human-explanation imitation.The comparison uses models trained on 15K explanations with different numbers of interpretable units K.Source: §4.3 · p. 8, Table 2 · p. 8
• Faithful human explanations increase Grad-CAM faithfulness by 7%, and adding the faithfulness loss produces a further 11% increase.The distribution of low faithfulness scores between 0.0 and 0.1 also decreases by over 17% under these changes.Source: §4.3 · p. 8, Table 3 · p. 8

• The approach generates textual explanations with words linked to relevant image regions actually attended to by the underlying VQA system.Source: §5 · p. 8
• Automated evaluations and crowdsourced human judgments demonstrate advantages over a previously published competing method.Source: §5 · p. 8
• Future work will integrate network dissection to include recognizable concepts from learned hidden-layer representations in explanations.Source: §5 · p. 9