Hello,
If I want to uses Params = Params + inverse(J'*J + mu * I) * g
instead of Params = Params + mu * g
,
where J
is a Jacobian matrix and J'
is transpose of J
, and J(i,j) = d loss(i) / d params(j)
,
and g
is a gradient vector, g = J' * loss
and I
is identity matrix,
and mu
is a scaler.
How can I compute jacobian matrix and inverse(J'*J + mu * I)
, and finally update parameters with the above formulation?
Thanks.