Complexity Control Facilitates Reasoning-Based Compositional Generalization in Transformers.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Transformers exhibit strong performance but their ability to handle compositional problems is debated.
• Understanding the internal mechanisms of Transformers is crucial for improving their generalization capabilities.

Purpose of the Study:

• To investigate the internal mechanisms of Transformers in compositional tasks.
• To identify factors influencing whether Transformers learn generalizable rules or rely on memorization.

Main Methods:

• Applied complexity control strategies (parameter initialization scale, weight decay).
• Utilized masking strategies on information circuits.
• Employed multiple complexity metrics to analyze internal workings.
• Validated findings across image generation and natural language processing datasets.

Main Results:

• Complexity control significantly influences solution type: reasoning-based (generalizable) vs. memory-based (memorized).
• Distinct internal mechanisms identified for reasoning-based and memory-based solutions.
• Reasoning-based solutions show a lower complexity bias, linked to neuron condensation.

Conclusions:

• Complexity control is key to enabling Transformers to learn reasoning rules for compositional tasks.
• Lower complexity bias in reasoning-based solutions is hypothesized to facilitate rule learning.
• Findings are broadly applicable across diverse real-world AI tasks.