From the paper (page 6 with a comparison to GLU and SwiGLU) they are not using tanh directly (i.e. f(x) = tanh(x)) but:
f_gate(b,x) = b * tanh(x / b) * sigmoid(x)
f_up(b,x) = b * tanh(x / b)
Looking at the graph I wonder if this is to try and get the best of both GLU (better representation at higher values of x >~ 5) and SwiGLU (the value bump just before 0).
From the paper (page 6 with a comparison to GLU and SwiGLU) they are not using tanh directly (i.e. f(x) = tanh(x)) but:
Looking at the graph I wonder if this is to try and get the best of both GLU (better representation at higher values of x >~ 5) and SwiGLU (the value bump just before 0).
Which part of that is doing gating? I thought gate generally looks like
for some f. In your case b is a constant hyperparameter though, e.g.
So where's the gate?
Still more of a tanh than I've seen in years
Wow this is fascinating enough that I’m actually going to read tfa lol