@lukem12345 pointed out that we can solve steady state problems with a backpropagation algorithm. One wrinkle is that we need to handle spans in the compute graph as below.
if the decapode has this pattern
a <-A- x -B->y
but
A\a != B\y
how do you set x during backward pass?
you can formulate
[a] = [A][x]
[b] = [B]
and solve it with:
[A;B][a;b]
@lukem12345 pointed out that we can solve steady state problems with a backpropagation algorithm. One wrinkle is that we need to handle spans in the compute graph as below.
if the decapode has this pattern
a <-A- x -B->y
but
A\a != B\y
how do you set x during backward pass?
you can formulate
[a] = [A][x]
[b] = [B]
and solve it with:
[A;B][a;b]