Source-linked AI summary
Towards Causal VQA: Revealing and Reducing Spurious Correlations by Invariant and Covariant Semantic Editing
Vedika Agarwal, Rakshith Shetty, Mario Fritz
TL;DR
VQA models can exploit spurious correlations rather than answer from image evidence, motivating robustness analysis beyond linguistic variation. The paper automatically creates invariant and covariant semantic image edits, measures prediction consistency, and uses the edited data for augmentation. Across diverse models and question types, including counting, the approach reduces flipping under semantic edits and improves robustness on synthetic and real data.
Problem
VQA models are brittle to linguistic and semantic variations and may rely on spurious correlations instead of image evidence.
Method
The paper automatically removes QA-irrelevant or QA-relevant objects to create invariant and covariant edits, then measures consistency and trains models with the synthetic data.
Results
The trained models show significantly less flipping under invariant and covariant semantic edits, with improvements on synthetic and real data.
Takeaways & Limitations
Semantic editing provides a way to reveal spurious correlations and improve VQA robustness toward visual variations.
Takeaways & Limitations
The edited robustness set selects IQAs with uniform ground-truth answers to avoid ambiguity, especially for counting and binary questions.
Abstract
from arXiv · showhide
Despite significant success in Visual Question Answering (VQA), VQA models have been shown to be notoriously brittle to linguistic variations in the questions. Due to deficiencies in models and datasets, today's models often rely on correlations rather than predictions that are causal w.r.t. data. In this paper, we propose a novel way to analyze and measure the robustness of the state of the art models w.r.t semantic visual variations as well as propose ways to make models more robust against spurious correlations. Our method performs automated semantic image manipulations and tests for consistency in model predictions to quantify the model robustness as well as generate synthetic data to counter these problems. We perform our analysis on three diverse, state of the art VQA models and diverse question types with a particular focus on challenging counting questions. In addition, we show that models can be made significantly more robust against inconsistent predictions using our edited data. Finally, we show that results also translate to real-world error cases of state of the art models, which results in improved overall performance.
1. Introduction
The paper introduces semantic image editing as a systematic way to expose spurious correlations in VQA and measure prediction consistency. It studies invariant and covariant edits, then uses edited data augmentation to improve robustness across models and question types.
- VQA models often rely on spurious correlations instead of image evidence and question intent, making predictions brittle to linguistic variations.
- Invariant edits remove objects irrelevant to the QA pair, while covariant edits remove mentioned objects and should produce predictable answer changes.
- Three diverse VQA models are analyzed, and the study emphasizes counting questions alongside broader question types.
- The paper proposes automatic semantic image manipulations, synthetic complementary data, and consistency metrics to quantify robustness to visual variations.
- Data augmentation with the edited set reduces inconsistent predictions and improves performance on both synthetic and real data.
2. Related Work
Prior VQA robustness work primarily studies linguistic modifications and answer-distribution changes. This paper instead focuses on robustness to semantic changes in images.
- Earlier VQA research spans recurrent, convolutional, attention-based, and compositional model architectures.
- Prior methods improve consistency by generating entailed or similar-intent questions and fine-tuning models when answers remain consistent.
3. Synthetic Dataset for Variances and Invariances in VQA
The paper constructs synthetic VQA datasets by automatically removing objects whose semantic relation to the question determines whether the answer should remain invariant or change. It uses vocabulary mapping, geometric safeguards, human validation, and complementary counting edits to create these variants.
- Dataset construction: The datasets are built from VQAv2 and MS-COCO using GAN-based re-synthesis to remove objects from 80 COCO categories.Object references in questions and answers are mapped to COCO categories before editing.
- Invariant and covariant edits: Objects absent from the question-answer pair are selected for invariant edits, while relevant objects are retained or removed to create answer-changing counting variants.Invariant edits should preserve the answer; covariant edits remove one counted instance so the answer changes from n to n −1.
- Editing constraints: Invariant object removal is allowed only when the largest instance occupies less than 10% of the image and its overlap score with QA-mentioned objects is below 10%.The overlap score uses dilated segmentation masks to avoid degrading the image or affecting relevant objects.
- Editing constraints: The edited invariant set includes only IQAs with uniform ground-truth answers to avoid ambiguity from multiple human responses.This restriction is especially relevant for counting and binary questions.
- Validation and scale: The fully automatic IV-VQA pipeline applies qualifying removals across 658k IQA triplets, and three human annotators agreed that edited IQAs were valid 91% of the time on a 4.96k subset.The validation subset was selected using inconsistency analysis of three VQA models.
- Invariant and covariant edits: For CV-VQA, the authors collect numeric counting questions, match target-object counts to human answers, remove one non-overlapping instance, and call the resulting answer-changing set covariant.The removed instance must occupy less than 10% of the image, and its overlap with other instances of the counted object must be zero.
4. Experiments: Consistency analysis
The experiments evaluate consistency under invariant and covariant semantic edits across three VQA models. Models frequently flip predictions, including on counting questions, revealing brittleness and reliance on spurious visual correlations.
- Consistency criteria: Invariant edits should preserve predictions, whereas covariant counting edits should reduce the answer from n to n-1.The datasets remove either irrelevant objects or one counted instance.
- Experimental setup: Three models—a CNN+LSTM, attention-based SAAA, and compositional SNMN—are evaluated for accuracy and consistency on original and edited VQA samples.Consistency is measured between real images and corresponding IV-VQA or CV-VQA edits.
- Invariant edits: On original VQA data, SAAA reaches 70.3% accuracy, compared with 66% for SNMN and 60.2% for CL.These accuracy figures are reported for the VQA v2 original data.
- Invariant edits: 6.9% neg→pos flips occur versus an expected 0.013% rate under irrelevant-object removal, indicating models use removed objects spuriously.The flip rate is interpreted as evidence that the removed object influenced predictions despite being irrelevant to the question-answer pair.
- Covariant edits: Counting models are inconsistent on more than 75% of CV-VQA edits, while original-set accuracy remains below 50% for all three models.SAAA achieves 49.9%, SNMN 47.9%, and CL 39.4% on the original counting set.
- Qualitative analysis: Qualitative examples show models changing answers after irrelevant-object removal or failing to count correctly after one counted instance is removed.These flips occur even when the original prediction is correct, so accuracy alone does not capture the models’ shortcomings.
5. Robustification by Data Augmentation
The paper uses invariant and covariant edited data to fine-tune VQA models and reduce inconsistent predictions. Synthetic data consistently lowers flipping across question types, with trade-offs in accuracy for counting-focused augmentation.
- Invariant augmentation: Fine-tuning combines real VQA data with IV-VQA edits, using question-type-specific subsets to test reductions in flipping relative to real-data fine-tuning.The comparison uses models fine-tuned on real data versus real plus invariant edits.
- Invariant augmentation: Synthetic data reduces flipping for every evaluated question type, although the reduction varies across models and question types.All plotted points lie above the y = 0 axis; the largest reductions differ between CL and SAAA.
- Invariant augmentation: Real+IV-VQA fine-tuning yields 5–6% relative improvement in flips across all models, with small or mixed accuracy changes.Flip values change from 17.15→16.1 for CL, 7.53→7.09 for SAAA, and 8.09→7.72 for SNMN.
- Covariant augmentation: CV-VQA augmentation reduces counting flips by 40% for all three models, accompanied by a 1–4% accuracy drop.CV-VQA edits remove one instance of the object being counted and target n/n-1 consistency.
- Covariant augmentation: Combining CV-VQA and IV-VQA data reduces counting flips by 30%, while specialized fine-tuning can also improve consistency and real-data accuracy.The paper reports these effects from the right-side counting experiments and qualitative fine-tuning examples.
6. Conclusion and Future Works
The paper concludes that semantic editing exposes spurious correlations in VQA models and that edited data can reduce inconsistent behavior. It positions these datasets as resources for progress toward causal VQA.
- Conclusion: Semantic editing provides an approach to study and quantify VQA robustness to visual variations.The approach explicitly evaluates prediction behavior under semantic image changes.
- Conclusion: The analyzed models are brittle to visual variations and exploit spurious correlations when predicting answers.This conclusion follows from inconsistent behavior under invariant and covariant edits.
- Conclusion: Training with edited data produces significantly less flipping under invariant and covariant semantic edits.The conclusion summarizes the reported robustification results.
- Future work: The synthetic data are made publicly available to support research toward causal VQA models.The paper describes the data as explicitly created for VQA robustness analysis and improvement.
A. Synthetic Dataset for Variances and Invariances in VQA
The consistency analysis uses a human-validated subset of edited IQAs whose labels flip for at least one of the three models.
- Validation subset: The validation subset contains 27,621 flipping IQAs, representing 25% of the IV-VQA test set.Flips may be positive or negative for any of the three evaluated models.
- Validation subset: Researchers sample up to 100 IQAs from each of 65 question categories, producing 4,960 edited IQAs for comparison.The sample is random within categories when enough examples are available.
- Validation subset: Answer flipping is treated as undesirable and forms the basis for robustness comparisons.The selected subset is intended to check that measured inconsistency is not caused by synthetic-dataset errors.
B.1. Models Training
The study compares three VQA architectures representing distinct design philosophies and trains them using specified online or official implementations.
- Model selection: Three models are selected: CNN+LSTM, attention-based SAAA, and compositional SNMN.CNN+LSTM uses ResNet152 image features and LSTM question features before classification.
- Training setup: The CNN+LSTM implementation is created by removing attention layers from the available SNMN code.
- Training setup: The models are trained with publicly available code and specified hyperparameters, while the validation split is reserved for testing.
B.2. Visualizations
Visualizations show that VQA models often change predictions after semantic image edits, including edits that should preserve or predictably change the answer. Attention maps do not clearly explain this behavior.
- IV-VQA predictions: Predictions from all three models are inconsistent between original and edited IV-VQA images.The evaluation expects consistent predictions across the paired images, but the visualized outputs frequently violate this expectation.
- CV-VQA predictions: Counting edits are expected to produce n/n-1 consistency, but enforcing this consistency appears to break the models completely.
- Flip-rate analysis: 7% versus 8.7%: pos→neg flip rates change only marginally between objects occupying 0-1% and 9-10% of image area.The analysis reports no large dependence between removed-object area and flip rate.
- Qualitative examples: The figures illustrate brittle responses to semantic image variations through prediction flips on IV-VQA and CV-VQA examples.
- Attention inspection: SAAA attention heatmaps are diffuse and do not clearly identify a single object receiving the model’s attention.Consequently, the visualized attention distributions are not conclusive for explaining prediction flips.
C. Robustification by Data Augmentation
Fine-tuning experiments evaluate robustness on a strict IV-VQA subset with zero overlap, and the resulting accuracy-flipping values remain comparable to those on the overall set.
- Evaluation subset: A strict IV-VQA subset with zero overlap is used for fine-tuning experiments and accuracy-flipping analysis.
- Evaluation subset: The strict-subset results are comparable to the models’ performance on the overall IV-VQA set.
C.2. InVariant VQA Augmentation
Synthetic edited data is used to fine-tune VQA models for greater consistency under invariant and covariant visual changes. The experiments report fewer flips and, in some cases, improved original-set accuracy.
- InVariant augmentation: Real-plus-synthetic fine-tuning always reduces flips, and sometimes also improves accuracy on the original VQA set.For CNN+LSTM on the ‘is this a’ question type, the reported reduction is (12.72-9.77)/12.72, approximately 23%.
- InVariant augmentation: Table 8 compares accuracy and flipping for models trained with real data against models trained with real and synthetic data.
- CoVariant augmentation: Table 9 presents the corresponding accuracy and flip results for CV-VQA augmentation.
- Consistency learning: The study argues that explicitly created original/edited pairs can support consistency losses across image-question-answer triplets.
- InVariant augmentation: Figure 10 visualizes examples where adding edited IV-VQA data makes model predictions more consistent.
- CoVariant augmentation: Figure 11 shows examples where synthetic data makes models n/n-1 consistent and accurate for covariant counting edits.