5 ms·
I am curious what's the actual formula. I mean, there so many headers and layers, it is tricky to make a choice that will resonate with our intuition . Is it s
by stared 8d ago
I am curious what's the actual formula.
I mean, there so many headers and layers, it is tricky to make a choice that will resonate with our intuition . Is it some weighted average? Or maybe ablation test?
- ifz 8d agoIt's really simple, basically just the magnitude of the value vector, weighted by QK dot product, summed across all attention heads and layers. When I started, I expected I'd have to experiment a lot to find something comprehensible. But this simple computation can already show some patterns.
- stared 8d agoNice! Sometimes the simplest approaches work the best.
- nullbio 7d agoMight be cool to try different colors for the different attention heads instead of summing them across all attention heads, and making the backgrounds composed of stacked color layers? So you can see how each attention head attends to tokens individually.
- visarga 8d agoIf you want quick access look at google images for "transformer attention formula" there are some interesting depictions