Unlock: Attention as Kernel Regression

Softmax attention viewed as Nadaraya-Watson kernel regression: the output at each position is a kernel-weighted average of values. Connects attention to classical nonparametric statistics and motivates linear attention via random feature approximations.