Multi-Head Attention as Ensemble Nadaraya-Watson Estimation: Variance Reduction, Decorrelation, and Optimal Head Diversity
Researchers have developed a statistical theory that frames multi-head attention (MHA) as an ensemble of Nadaraya-Watson kernel regression estimators. This framework reveals that variance reduction in MHA is fundamentally tied to the decorrelation of outputs from different attention heads, rather than just the number of heads. They introduced the Head Diversity Index (HDI) to measure this decorrelation and derived an optimal head-dimension allocation strategy, suggesting a new architectural scaling law where optimal per-head dimension grows logarithmically with training set size. AI
IMPACT Provides a theoretical basis for understanding and optimizing attention mechanisms in large language models.