Behind the Scene: Revealing the Secrets of Pre-trained Vision-and-Language Models (2005.07310v2)

Abstract: Recent Transformer-based large-scale pre-trained models have revolutionized vision-and-language (V+L) research. Models such as ViLBERT, LXMERT and UNITER have significantly lifted state of the art across a wide range of V+L benchmarks with joint image-text pre-training. However, little is known about the inner mechanisms that destine their impressive success. To reveal the secrets behind the scene of these powerful models, we present VALUE (Vision-And-Language Understanding Evaluation), a set of meticulously designed probing tasks (e.g., Visual Coreference Resolution, Visual Relation Detection, Linguistic Probing Tasks) generalizable to standard pre-trained V+L models, aiming to decipher the inner workings of multimodal pre-training (e.g., the implicit knowledge garnered in individual attention heads, the inherent cross-modal alignment learned through contextualized multimodal embeddings). Through extensive analysis of each archetypal model architecture via these probing tasks, our key observations are: (i) Pre-trained models exhibit a propensity for attending over text rather than images during inference. (ii) There exists a subset of attention heads that are tailored for capturing cross-modal interactions. (iii) Learned attention matrix in pre-trained models demonstrates patterns coherent with the latent alignment between image regions and textual words. (iv) Plotted attention patterns reveal visually-interpretable relations among image regions. (v) Pure linguistic knowledge is also effectively encoded in the attention heads. These are valuable insights serving to guide future work towards designing better model architecture and objectives for multimodal pre-training.

Insights into Pre-trained Vision-and-LLMs

The paper, "Behind the Scene: Revealing the Secrets of Pre-trained Vision-and-LLMs," provides a comprehensive investigation into the inner workings of Transformer-based vision-and-language (V+L) models. With models like ViLBERT, LXMERT, and UNITER achieving state-of-the-art performance across various benchmarks, understanding the mechanisms of their success is crucial for advancing multimodal research. The paper introduces VALUE (Vision-And-Language Understanding Evaluation), a series of probing tasks designed to decode the implicit knowledge ingrained in the attention mechanisms and contextual embeddings within these models.

Key Findings

The paper focuses on evaluating two primary architectures: single-stream (UNITER) and two-stream (LXMERT) models. Through meticulous experiments, several key observations were made:

1. Multimodal Fusion Dynamics: In single-stream models, deeper layers exhibit a propensity for intertwining multimodal fusion, leading to increasingly indistinguishable representations of text and image modalities. Contrarily, the two-stream model showcases the opposite trend where fusion becomes less pronounced as it progresses through layers.
2. Modality Influence: Analysis of attention traces reveals that pre-trained models exhibit a bias towards textual modality during prediction, with text exerting a more dominant influence over their decision-making process compared to visual inputs.
3. Cross-modal Interaction: A subset of attention heads is identified as specialized for cross-modal interaction, with these heads effectively capturing alignment and semantic links between visual regions and textual elements. This capability develops organically in single-stream models but is structured by design in two-stream architectures.
4. Visual Relations: The pre-trained models encapsulate significant knowledge regarding visual relations between image regions, which is demonstrated by their performance in visual relation detection tasks.
5. Linguistic Encoding: Despite their focus on multimodal pre-training, the models also encode rich linguistic knowledge. Single-stream models, particularly those initialized with BERT weights, exhibit superior performance in language understanding tasks compared to two-stream models.

Implications and Future Directions

The implications of these findings are multifold. On a theoretical level, they elucidate the nature of multimodal fusion and the role of attention heads in encoding cross-modal insights. Practically, understanding these mechanisms can guide the design of more effective V+L architectures and improve task-specific applications such as image-text retrieval, visual question answering, and referring expression comprehension.

Several recommendations arise from this research for future model development and analysis:

• Model Design: Continued exploration of single-stream architectures is warranted given their capacity to capture comprehensive intra- and cross-modal information. Future models may benefit from single-stream designs that incorporate BERT-like initialization for enhanced language understanding.
• Evaluation Frameworks: The probing tasks described offer a framework for evaluating intermediate checkpoints during pre-training, allowing researchers to assess learned knowledge before fine-tuning on downstream tasks, thus optimizing training efficiency.
• Attention Supervision: Incorporating explicit supervision tied to probing tasks could further refine model training, leading to systems that are not only performant but also interpretable.

The paper sets a precedent for detailed analysis of pre-trained models in V+L research, providing valuable insights that could inform the evolution of multimodal AI systems. With potential future work focusing on compression and pruning based on probe results, this paper represents a significant contribution in unpacking the complex dynamics of attention-driven multimodal model behavior.