Reproduce “Functional Equivalence in Attention: A Comprehensive Study with Applications to Linear Mode Connectivity” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Viet-Hoang Tran, Vinh Khanh Bui, Van-Hoan Trinh, Tan Lai Ngoc, Tan M. Nguyen
This research formally characterizes the functional equivalence of Transformer attention mechanisms incorporating positional encodings, showing that Rotary Positional Encoding (RoPE) reduces architectural symmetry, which provides a theoretical basis for its enhanced expressivity. It also introduces a two-stage weight-matching algorithm to empirically demonstrate linear mode connectivity (LMC) in various Transformers, revealing LMC's dependence on model scale and architecture, with connected models exhibiting improved robustness to distribution shifts.
Key insights
- Rotary Positional Encoding (RoPE) formally reduces the symmetry group of Multihead Attention, offering a theoretical explanation for its enhanced expressivity and practical success.
- Linear Mode Connectivity (LMC) in Transformers is not universal, showing variability based on the type of positional encoding, model architecture, and the extent of parameter re-initialization.
- Proper alignment of attention heads and internal parameters through a specialized matching algorithm is critical for uncovering underlying LMC.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.