PaperlayerView exact revision on arXiv

Source-linked AI summary

Taking a HINT: Leveraging Explanations to Make Vision and Language Models More Grounded

Ramprasaath R. Selvaraju, Stefan Lee, Yilin Shen, Hongxia Jin, Shalini Ghosh, Larry Heck, Dhruv Batra, Devi Parikh

arXiv:1902.03751v2cs.CV

TL;DR

Vision-and-language models can rely on language priors instead of relevant visual evidence, weakening their grounding. HINT aligns gradient-based network importances with human attention to make models rely on human-relevant regions. Applied to VQA and image captioning, it improves grounding and reaches an 8-percentage-point VQA-CP state-of-the-art gain with only 6% human supervision.

  • Problem

    Vision-and-language models often rely on superficial linguistic correlations rather than appropriately associating words and phrases with relevant image regions.

  • Method

    HINT uses a ranking loss and gradient-of-gradient updates to align gradient-based network explanations with human importance annotations.

  • Results

    8 percentage points on VQA-CP with human importance supervision for 6% of training data, alongside improved visual grounding in VQA and image captioning.

  • Takeaways & Limitations

    Better grounding improves generalization to changing test distributions and increases perceived model trustworthiness.

  • Takeaways & Limitations

    Network importance is computed for the ground-truth decision rather than the predicted decision because human attention for incorrect decisions is unavailable and conceptually ill-posed.

Abstract

from arXiv · show

Many vision and language models suffer from poor visual grounding - often falling back on easy-to-learn language priors rather than basing their decisions on visual concepts in the image. In this work, we propose a generic approach called Human Importance-aware Network Tuning (HINT) that effectively leverages human demonstrations to improve visual grounding. HINT encourages deep networks to be sensitive to the same input regions as humans. Our approach optimizes the alignment between human attention maps and gradient-based network importances - ensuring that models learn not just to look at but rather rely on visual concepts that humans found relevant for a task when making predictions. We apply HINT to Visual Question Answering and Image Captioning tasks, outperforming top approaches on splits that penalize over-reliance on language priors (VQA-CP and robust captioning) using human attention demonstrations for just 6% of the training data.

1. Introduction

Vision-and-language models often rely on linguistic correlations instead of grounding decisions in relevant image regions. HINT addresses this by aligning gradient-based explanations with human attention, improving grounding across VQA and image captioning.

  • Many vision-and-language models associate words poorly with image regions and rely on superficial linguistic correlations.Their performance drops substantially on datasets designed to penalize language-prior reliance.
  • Gradient-based analyses show that models often focus on irrelevant regions, such as a produce stand instead of bananas.These regions can differ substantially from those humans select for the same task.
  • Standard input-output training offers no explicit grounding guidance, leaving models to resolve visual associations from co-occurrences.Language tokens can provide easier correlations than learning visual variation and shared visual features.
  • Explicitly supervising model attention can improve attention-map similarity without reducing language bias.Because attention feeds later classifiers, models may attend to appropriate regions while still ignoring their visual content.
  • HINT aligns gradient-based explanations with human importance annotations using a ranking loss and gradient-of-gradient parameter updates.This constrains models to be sensitive to relevant visual content, not merely to look at the correct regions.
  • With human importance supervision for 6% of training data, HINT improves VQA-CP state of the art by 8 percentage points and improves grounding in VQA and captioning.Human studies also find HINT-trained models more reasonable and trustworthy.

2. Related Work

Related work addresses interpretability, bias-sensitive evaluation, debiasing, and human attention supervision in vision-and-language models. HINT differs by aligning gradient-based importances with human attention to improve grounding rather than only model attention.

  • Interpretability research develops explanations from internal states and gradient-based importance measures for deep networks.HINT builds on this explanatory perspective while targeting visual grounding.
  • VQA-CP changes answer distributions between training and test, exposing models that memorize language associations instead of grounding answers in images.A robust captioning split similarly tests performance under bias-sensitive conditions.
  • Prior debiasing methods target gender-word generation, disentangle visual and language components, or use adversarial regularization.GVQA can be effective but requires multi-stage training and is difficult to extend to new models.
  • Human attention maps for VQA are collected by having people deblur image regions until they can confidently answer questions.HINT uses these maps to supervise gradient-based visual explanations.
  • Supervising model attention improves attention correlation but does not necessarily reduce language bias because later layers can disregard visual signals.Gradient explanations are presented as more directly linked to model decisions and input regions.
  • HINT contrasts with related gradient-importance work by aligning importance scores with human attention maps to improve network grounding.The comparison distinguishes grounding-oriented alignment from learning classifiers using domain knowledge.

3. Preliminaries

The paper uses Bottom-Up Top-Down attention architectures as its base for VQA and image captioning. These models attend over salient region proposals using question or language-state information before producing predictions.

  • The model-agnostic HINT experiments use the Bottom-Up Top-Down architecture as the base model.The architecture uses top-down attention for fine-grained and multi-stage vision-language reasoning.
  • Bottom-Up Top-Down Attention for VQA: For VQA, UpDown encodes up to k salient image-region proposals with Faster-RCNN and computes question-guided soft attention over them.The pooled attended feature is combined with the question feature for prediction.
  • Bottom-Up Top-Down Attention for Image Captioning: The image-captioning model contains attention and language LSTM networks that process visual features and generated-word context.The attention LSTM selects proposal features, while the language LSTM predicts the next vocabulary word.

4. Human Importance-aware Network Tuning

HINT aligns gradient-based network importance with human attention over image regions, training models to rely on visually relevant content while preserving task performance. It applies this alignment to VQA and image captioning using limited human supervision.

  • Human Importance-aware Network Tuning: HINT estimates input-region importance with gradient-based explanations and tunes network parameters to align those scores with human-designated important regions.The method updates parameters through a gradient-of-gradient step.
  • Human Importance: Human importance scores are computed for region proposals from normalized attention-map energy inside versus outside each proposal.For VQA, human attention maps vary across questions for the same image.
  • Human Importance: Human attention supervision covers approximately 6% of the VQA dataset, with 40,554 image-question pairs available.For image captioning, COCO segmentation annotations are mapped to visual words and used consistently on a 6% subset.
  • Network Importance: Network importance is computed for the ground-truth answer or visual word by differentiating its score with respect to proposal features.The method uses the ground-truth decision rather than the predicted decision because human attention for incorrect decisions is unavailable and ill-posed.
  • Human-Network Importance Alignment: A ranking loss penalizes proposal pairs whose network-importance order disagrees with human importance, using relative rankings because the score scales differ.The loss considers all misranked proposal pairs and updates with their network-importance differences.
  • Human-Network Importance Alignment: The final objective combines the HINT ranking loss with the original task loss so models align importance while retaining VQA or captioning performance.The first term encourages reliance on correct regions, while the task term encourages correct predictions; optimization requires second-order gradients.

5. Experiments and Analysis

Experiments on VQA and image captioning show that HINT improves visual grounding, especially on bias-sensitive splits, while attention alignment does not. The gains require limited human supervision and can improve task performance without reducing standard VQA accuracy.

  • VQA: HINT improves standard VQA performance while also improving the compositional VQA-CP split.The paper contrasts this with prior approaches reporting trade-offs between the two splits.
  • VQA: Direct attention alignment fails to reduce language bias on VQA-CP, whereas HINT’s gradient supervision improves grounding.The authors attribute the difference to gradient-based explanations affecting all network parameters and more faithfully reflecting decisions.
  • VQA: With human attention supervision for 1.5% of the VQA dataset, HINT achieves a 5% performance improvement.Using all available human attention maps corresponds to 6% of VQAv2 data.
  • Image Captioning: On robust captioning, HINT raises CIDEr from 0.90 to 0.92, while standard-split training decreases CIDEr from 1.06 to 1.04.The grounding improvement is therefore not fully captured by captioning score changes.
  • Image Captioning: Qualitative captioning examples show improved localization of visual words and individual object occurrences, including fork, apple, orange, and a cat shadow.The VQA examples likewise show HINT attending to relevant objects such as a boy and skateboard.

6. Evaluating Grounding

Direct grounding evaluation compares gradient-based network importance with human attention. HINT increases their rank correlation on both VQA and image captioning, and the captioning study also evaluates human trust.

  • Grounding Evaluation: Grounding is evaluated by comparing network importance for ground-truth decisions with human attention and computing Spearman rank correlation.The evaluation uses VQA-CP and COCO robust test images.
  • Human Trust: An AMT study compares baseline and HINTed captioning explanations and asks workers to rate relative reasonableness on a five-point Likert scale.The interface presents a ground-truth visual word, its explanation, and the whole caption.
  • VQA: On VQA, HINT raises Spearman correlation with human attention from -0.09 to 0.18.The paper notes that values near 0 indicate no correlation on the -1 to 1 scale.
  • Image Captioning: On image captioning, HINT raises rank correlation with COCO segmentation maps from 0.008 to 0.17.The evaluation uses visual-word explanations on the COCO robust split.
  • Grounding Evaluation: The improved visual grounding generalizes to new images and language contexts better than the baseline model.This conclusion follows from the rank-correlation evaluation on held-out test splits.

7. Evaluating Trust

The study tests whether improved grounding increases human trust in captioning models while controlling for prediction accuracy. Participants compared explanations from base and HINTed models using ground-truth captions.

  • Participants compared base and HINTed captioning models using explanations for ground-truth visual words and captions.Both models were shown predictions corresponding to the ground-truth caption, isolating grounding from accuracy differences.
  • 42 AMT workers produced 1000 responses covering 200 image pairs in the trust study.Each image pair received five annotations.
  • 49.9% of instances preferred HINT compared with 33.1% preferring the base model.Participants rated the relative reasonableness of the two models on a five-point Likert scale.

8. Does HINT also improve model attention?

HINT improves feed-forward model attention as well as gradient-based grounding. These attention changes are associated with higher trust judgments for HINTed VQA-CP models.

  • 0.57 IoU for UpDn on VQA-CP increased to 0.63 after applying HINT.The metric measures overlap between the top-scoring proposal box and human attention maps.
  • 31% of respondents found the HINTed VQA-CP model more trustworthy versus 16.5% for the base model.The study collected 10 responses for each of 100 randomly sampled image-question pairs.

9. Conclusion

HINT aligns network sensitivity with human-identified relevant image regions and improves visual grounding in VQA and image captioning. The paper also connects better grounding with generalization to changing test distributions and model trustworthiness.

  • HINT aligns network sensitivity to spatial input regions that humans deem relevant to a task.The framework is presented as a general method for improving visual grounding in vision-and-language tasks.
  • HINT improves visual grounding in both Visual Question Answering and image captioning.
  • Regularizing network gradients for grounding may apply beyond vision and language by enabling focused feedback to networks.

A. Qualitative examples

Qualitative examples show HINTed VQA and captioning models attending to more appropriate regions and sometimes answering correctly where base models fail. Captioning examples also show more precise localization of visual words and individual object occurrences.

  • VQA examples: HINTed VQA models answer “yes” to the screaming question and localize both the person’s face and the bear, unlike the incorrect base model.
  • VQA examples: For the clock question, HINT changes the answer from the base model’s incorrect “no” to “yes” and localizes the clock.
  • VQA examples: HINT improves localization even when both VQA models produce incorrect answers.
  • Captioning examples: HINTed captioning models attend more accurately to faces, ties, spoons, and sinks when generating corresponding visual words.
Loading 1902.03751v2…