How Transferable Are Reasoning Patterns in VQA?

Corentin Kervadec, Theo Jaunet, Grigory Antipov, Moez Baccouche, Romain Vuillemot, Christian Wolf

Abstract

Since its inception, Visual Question Answering (VQA) is notoriously known as a task, where models are prone to exploit biases in datasets to find shortcuts instead of performing high-level reasoning. Classical methods address this by removing biases from training data, or adding branches to models to detect and remove biases. In this paper, we argue that uncertainty in vision is a dominating factor preventing the successful learning of reasoning in vision and language problems. We train a visual oracle and in a large scale study provide experimental evidence that it is much less prone to exploiting spurious dataset biases compared to standard models. We propose to study the attention mechanisms at work in the visual oracle and compare them with a SOTA Transformer-based model. We provide an in-depth analysis and visualizations of reasoning patterns obtained with an online visualization tool which we make publicly available (https://reasoningpatterns.github.io). We exploit these insights by transferring reasoning patterns from the oracle to a SOTA Transformer-based VQA model taking standard noisy visual inputs via fine-tuning. In experiments we report higher overall accuracy, as well as accuracy on infrequent answers for each question type, which provides evidence for improved generalization and a decrease of the dependency on dataset biases.

Introduction

Despite high performance on benchmarks, VQA models are notorious for "Clever Hans" effects—exploiting statistical shortcuts in datasets rather than performing genuine reasoning. For instance, if a dataset frequently features yellow bananas, a model might learn to answer "yellow" to any color-related question about bananas without actually inspecting the image.

This research argues that the primary bottleneck is visual uncertainty. When object detectors are noisy or miss objects entirely, models are forced to rely on language biases to guess answers. The core challenge is teaching a model to perform the "algebraic manipulation of words and objects" (reasoning) when the visual inputs are blurry, mislabeled, or missing.

The Oracle Transfer Method

To decouple reasoning from the noise of visual perception, the authors propose a strategy called Oracle Transfer. This method involves training a "Visual Oracle"—a Transformer-based architecture (tiny-LXMERT)—using Ground Truth (GT) visual information. This provides the model with perfect object classes and bounding boxes instead of noisy features from a standard detector like Faster R-CNN.

Overview of the Oracle Transfer pipeline, showing the transition from a Ground Truth (GT) trained Oracle to a fine-tuned model using noisy R-CNN features.

Once the Oracle is trained, its weights are used to initialize a standard "deployable" model that uses noisy R-CNN features. This model is then fine-tuned on real, noisy data, allowing it to retain the high-level reasoning patterns learned by the Oracle while adapting to real-world visual uncertainty.

Analyzing Attention Modes

To understand how the Oracle "thinks," the researchers analyzed its attention heads and categorized them into "modes" based on their energy distribution:

  • Dirac: Highly focused on a single token.
  • Uniform: Spread evenly across all tokens.
  • Bimorph: A hybrid mode where the head switches between being focused or uniform depending on the specific question type (e.g., "choose color" vs. "verify size").
Example of a Bimorph attention mode, characterized by a hybrid distribution that adapts to the reasoning task.

Experimental Results

The method was evaluated on the GQA dataset, the GQA-OOD (Out-Of-Distribution) benchmark, and VQAv2. Key findings include:

  • Generalization Boost: On GQA-OOD, Oracle Transfer improved accuracy on rare answers ("acc-tail") from 42.9% to 48.5%, indicating a significant reduction in bias dependency.
  • Overall Accuracy: On the standard GQA test-dev set, the method improved overall accuracy from 52.4% to 56.8%.
  • Cross-Dataset Success: Transferring reasoning patterns from a GQA-Oracle to a model trained on VQAv2 resulted in a gain of +0.5 points, showing that these patterns are not just specific to a single dataset's templates.

Key Insights and Conclusion

The study demonstrates that vision is the primary bottleneck for reasoning; given perfect sight, models naturally rely less on dataset biases. Furthermore, reasoning appears to be modular, with specific attention heads specializing in functions like identifying colors or verifying sizes.

Most importantly, the pattern of attention—the strategy of which objects to relate to which words—is transferable. Even when the underlying visual features change from perfect symbols to noisy pixels, the reasoning strategy remains valid. Future work will explore how these patterns hold up in unconstrained natural language environments and seek to further interpret the symbolic nature of these transferred patterns.

Summary updated

Abstract

Since its inception, Visual Question Answering (VQA) is notoriously known as a task, where models are prone to exploit biases in datasets to find shortcuts instead of performing high-level reasoning. Classical methods address this by removing biases from training data, or adding branches to models to detect and remove biases. In this paper, we argue that uncertainty in vision is a dominating factor preventing the successful learning of reasoning in vision and language problems. We train a visual oracle and in a large scale study provide experimental evidence that it is much less prone to exploiting spurious dataset biases compared to standard models. We propose to study the attention mechanisms at work in the visual oracle and compare them with a SOTA Transformer-based model. We provide an in-depth analysis and visualizations of reasoning patterns obtained with an online visualization tool which we make publicly available (https://reasoningpatterns.github.io). We exploit these insights by transferring reasoning patterns from the oracle to a SOTA Transformer-based VQA model taking standard noisy visual inputs via fine-tuning. In experiments we report higher overall accuracy, as well as accuracy on infrequent answers for each question type, which provides evidence for improved generalization and a decrease of the dependency on dataset biases.
BibTeX
@inproceedings{kervadec2021how,
  author = {Corentin Kervadec and Theo Jaunet and Grigory Antipov and Moez Baccouche and Romain Vuillemot and Christian Wolf},
  title = {{How Transferable Are Reasoning Patterns in VQA?}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2021},
  url = {https://openreview.net/forum?id=PKD_c3781Gx},
}
Figure