Explicit Attention Matrices for Interpretability and Custom Masks

you still materialise the matrix when you need to look at it or apply a mask no fused kernel supports — the naive form is a tool, not a mistake

This Concept is waiting for its first lesson!

you still materialise the matrix when you need to look at it or apply a mask no fused kernel supports — the naive form is a tool, not a mistake

Are you a teacher? Sign in to start contributing.