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. 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 in the dependency on dataset biases.

Introduction: The Vision Bottleneck

Visual Question Answering (VQA) models are notorious for being "Clever Hans" agents: they often provide the right answer for the wrong reasons. Instead of reasoning about the objects in an image (e.g., "Is the person wearing a red hat?"), they exploit statistical shortcuts (e.g., "Most hats in this dataset are red, so I'll just guess red").

This research argues that messy, real-world visual input is the primary driver of this behavior. When an object detector misses a key item or mislabels a "fork" as a "knife," the reasoning engine abandons logic in favor of dataset biases. This paper investigates whether we can "teach" a model logic using perfect visual data and then transfer that reasoning to a model dealing with noisy, real-world images.

Methodology: The Oracle Transfer

The researchers employed a three-step strategy to bridge the gap between perfect and noisy vision:

  1. The Visual Oracle: A "perfect-sighted" model was built by feeding it ground-truth (GT) object labels and coordinates instead of raw image features. This isolated the logical relationship between words and objects.
  2. Attention Analysis: The team analyzed the "attention heads" of the Transformer-based model using a metric called the $k$-distribution—the number of tokens a model needs to reach 90% of its "attention energy." They identified three distinct modes: Dirac (focused on one object), Uniform (broad scene attention), and Bimorph (dynamic switching between focused and broad).
  3. Oracle Transfer: A standard VQA model (using noisy R-CNN features) was initialized with the weights of the Visual Oracle and then fine-tuned on real data. This effectively transferred the reasoning patterns learned from perfect data to a model with imperfect vision.
The Oracle Transfer pipeline: training on ground-truth data to learn reasoning patterns, then fine-tuning on noisy data.

Key Insights: Reasoning vs. Bias

The study reveals that the "reasoning" components of deep learning models are more capable than previously thought; the uncertainty of visual input is the true bottleneck. When vision is perfect, the model naturally ignores biases.

In the Oracle model, specific attention heads specialized in tasks like "filtering by color" or "verifying size." While these heads are "muddier" in noisy models, the transfer process helps preserve this modularity. Furthermore, the most effective reasoning heads were found to be "Bimorph," dynamically switching between a "search" mode and a "found it" mode based on the question context.

Accuracy vs. question/answer frequency: The perfect-sighted Oracle is significantly more robust to infrequent (tail) data compared to models with visual noise.

Experimental Results

The proposed Oracle Transfer method was evaluated on several benchmarks, including GQA, GQA-OOD (Out-Of-Distribution), and VQAv2.

  • Generalization Boost: On the GQA-OOD "tail" (the rarest, hardest samples), the Oracle Transfer method improved accuracy from 42.9% to 48.5%, demonstrating a significant reduction in bias dependency.
  • Efficiency: A relatively small model (26M parameters) achieved accuracy competitive with much larger state-of-the-art models like MCAN-6 (52M parameters) while maintaining superior robustness.
  • Qualitative Success: Visualizations showed that the transferred model could overcome visual errors. For instance, if a detector mislabeled a "fork," the transferred model's attention was strong enough to still "look" at the correct object to answer the question, whereas baseline models failed.

Conclusion and Future Work

This work demonstrates that reasoning patterns are indeed transferable from a perfect-sighted oracle to a standard noisy model. By pre-training "logic" on ground-truth data, models can be made more robust to the inherent uncertainty of computer vision.

Future research will address the "first layer" problem—the difficulty of transferring the initial visual layers due to differing input formats—and explore how these reasoning patterns hold up in unconstrained, "in-the-wild" video streams beyond synthetic benchmarks like GQA.

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