OONumerics User : |
From: Georg Holzmann (grh_at_[hidden])
Date: 2007-07-23 04:43:14
Hallo Michael!
Thanks for the detailed answer!
> - If you really want an efficient solution you also have to consider
> things like: do I need all eigenvalues, or only the
> five largest eigenvalues?
Actually I only need the largest eigenvalue.
The basic algorithm (which is an "echo state network") can be divided in
three parts:
1) network generation (not performance critical - is made in advance):
here I need eigenvalue calculation, maybe SVD, some scaling of the
sparse matrix and random values etc.
2) simulation of the network (performance critical): basic arithmetic
functions (multiplication with vector, addition), applying some
activation-function (e.g. tanh) to all the values of the sparse matrix,
random values
3) training:
- offline version (not performace critical): a pseudo inverse has to be
calculated (to solve the least square problem), which should result
again in a sparse matrix
- online version (performace critical): instead of the pseudo inverse an
LMS or/and RLS algorithm will be implemented
Thanks for any hint,
LG
Georg