|
I am trying to apply the method shown in this paper: Schraudolph, N. N. (2002). Fast curvature matrix-vector products for second-order gradient descent. Neural computation, 14(7), 1723–38, to a simple model, but I'm not sure of my implementation. Apologies for the cross-post, but the math markup is much better on the math.stackexchange. The notation is here. I am trying to understand this technique better and also check my work. |
|
Hi, James Martens has implemented perfectly Schraudolph's method for all the Hessian, Gauss-Newton and Fisher information matrices, within Matlab. You need to take a look at his code at: http://www.cs.toronto.edu/~jmartens/docs/HFDemo.zip Cheers, -- Arya Thank you, I will do that!
(Dec 13 '12 at 19:11)
Neil
Hmm, the code is not in a great state right now. There are thousands of lines of dead code…
(Dec 13 '12 at 21:05)
Neil
Neil - have you had any luck ?
(Jan 04 '13 at 13:45)
will henry
|
