Abstract
This paper investigates whether state-of-the-art Vision-Language (VL) models like UNITER, LXMERT, and ViLT truly integrate information from both images and text. By employing a specialized probing framework and adversarial datasets, the authors reveal that while these models excel at multimodal color identification, they frequently rely on linguistic biases for spatial reasoning (size and position). Furthermore, the study highlights the phenomenon of "catastrophic forgetting" of language skills during the fine-tuning process.
Introduction
Transformer-based architectures have become the standard for Vision-Language tasks such as Visual Question Answering (VQA). Despite their high performance, it remains unclear if these models develop genuine multimodal representations or simply exploit statistical shortcuts within the training data. For instance, a model might predict "green" for "grass" based solely on textual frequency rather than visual evidence. This research aims to "probe" the internal representations of these models to understand what information is actually encoded and how it evolves from pre-training to fine-tuning.
Methodology: The Probing Framework
The researchers utilize a probing framework to analyze the internal states of VL models. Instead of evaluating final task outputs, they extract internal representations (vectors from the final layer) and train simple linear probes to perform specific diagnostic tasks.
A critical innovation is the introduction of mismatched baselines (denoted as $p^♣$). To distinguish between true multimodal integration and reliance on a single modality (e.g., text bias), the probe is tested on examples where the image and caption are intentionally mismatched. If the probe's performance remains high despite the mismatch, it indicates the model is relying on unimodal shortcuts rather than cross-modal reasoning.
The study evaluates three major architectures:
- UNITER & LXMERT: Region-based models using a Faster R-CNN backbone to extract object regions.
- ViLT: A patch-based, "convolution-free" model that processes raw image patches directly.
Datasets and Tasks
The evaluation spans several dimensions using both existing and newly created datasets:
- Language Probing: Subsets of Flickr30k for Part-of-Speech tagging and Bigram Shift detection.
- Vision Probing: Flower-102 for fine-grained classification and MS-COCO for object counting.
- Multimodal Probing: Four new datasets designed to minimize bias:
- M-Col: Color identification.
- M-Size & M-Pos: Size and position detection, balanced to prevent reliance on object-category priors (e.g., "elephants are big").
- M-Adv: Adversarial captions where key words are replaced with visually incorrect but linguistically plausible alternatives using BERT.
Key Findings
Color vs. Spatial Reasoning
Models demonstrate a genuine ability to link colors in text to visual features. In the M-Col task, UNITER achieved 86.27% accuracy, which dropped to 34.80% with mismatched images, confirming its reliance on visual input. Conversely, for size and position (M-Size/M-Pos), performance often hovered near the majority or mismatched baselines, suggesting these spatial concepts are poorly integrated at a multimodal level.
The Visual Bottleneck
The "frozen" object detectors used in UNITER and LXMERT appear to be a bottleneck. ViLT, which processes raw patches, significantly outperformed region-based models on vision-only tasks, achieving 91.34% on Flower-102 compared to UNITER's 71.82% and LXMERT's 75.56%.
Catastrophic Forgetting
Fine-tuning on specific tasks like VQA can lead to a loss of general linguistic capabilities. LXMERT, for example, showed a marked decrease in language understanding after fine-tuning, highlighting a trade-off between task-specific optimization and general representation quality.
Conclusion
The probing perspective reveals that while VL Transformers are making strides in multimodal integration, they remain heavily influenced by linguistic biases, particularly for spatial reasoning. The results suggest that moving away from fixed object detectors toward more flexible visual front-ends and addressing "forgetting" during fine-tuning are essential steps toward building truly multimodal systems.
Summary updated