Representational Capacity: Geometric Limits on Feature Representation in Transformer Language Models
Researchers have developed a new framework to understand the geometric limits of feature representation in transformer language models. By analyzing the embedding matrix and its deviation from near-orthogonality, they identified two classes of models: those with high deviation lacking structure and those with low deviation maintaining it. This work refines the understanding of representational capacity, showing it's exponentially sensitive to orthogonality constraints and that larger models prioritize tighter constraints over maximizing raw capacity. AI
IMPACT Provides a theoretical lens for understanding model limitations and potential improvements in feature representation.