Attention Flows: Analyzing and Comparing Attention Mechanisms in Language Models.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Computer Vision

Background:

• Deep attention-based models are state-of-the-art for Natural Language Processing (NLP) tasks.
• These models undergo pre-training on large datasets and fine-tuning for specific applications.
• Understanding how fine-tuning affects attention mechanisms in these models remains a challenge.

Purpose of the Study:

• To propose a visual analytics approach for understanding fine-tuning in attention-based language models.
• To introduce a visualization tool, Attention Flows, for querying, tracing, and comparing attention patterns.
• To provide insights into how classification decisions are made by visualizing attention flow.

Main Methods:

• Developed a visualization tool named Attention Flows.
• Designed the visualization to focus on classification-based attention in the deepest layer.
• Incorporated features for tracing attention flow from prior layers to input words.
• Enabled comparison between pre-trained and fine-tuned models.

Main Results:

• Attention Flows facilitates the analysis of attention mechanisms within and across layers and heads.
• The tool supports visual comparison of attention patterns between pre-trained and fine-tuned models.
• Demonstrated the evolution of attention mechanisms for various sentence understanding tasks.

Conclusions:

• Attention Flows provides valuable insights into the fine-tuning process of attention-based language models.
• The visualization aids in understanding task-specific attention adaptations.
• This approach enhances the interpretability of complex language models.