b 1 0 = 0 , b 2 0 = 0
For r = 1 to T do
c 1 r , c 2 r = softmax ( b 1 r − 1 , b 2 r − 1 )
s r = c 1 r u 1 + c 2 r u 2
a r = S q u a s h ( s r )
b i r = b i r − 1 + a r • u i