Use broadcasting to construct the Kronecker's delta

src/back.jl

@@ -173,11 +173,8 @@ Calculate the output jacobian `J = d/dx m(x)` such that each row `i` of `J` corr
 """
 function jacobian(f, x::AbstractVector)
   y::AbstractVector, back = forward(f, x)
-  function ȳ(i)
-    δ = fill!(similar(y, Bool), false)
-    δ[i] = true
-    return δ
-  end
+  # Using broadcasting so that output of `ȳ` is a GPU array if `y` is so:
+  ȳ(i) = ((j, _) -> i == j).(1:length(y), y)
   vcat([transpose(back(ȳ(i))[1]) for i = 1:length(y)]...)
 end