Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: More information about `annfit'?

From   Stas Kolenikov <>
Subject   Re: st: More information about `annfit'?
Date   Tue, 19 Apr 2011 10:00:23 -0500

Na-ah, it's wa-a-ay simpler than that. It is a random search over the
parameter space (as the objective function is multimodal, gradient
optimization will make little sense), and it probably relies on
logistic ogive. It was written under Stata 5 or 6, so if you started
programming Stata with version 11, the code will look very archaic,
but it should still be readable. Programming back-propagation was a
fair chunk of coding work that I did not have the time or need to
perform back then.

On Tue, Apr 19, 2011 at 8:49 AM, Diego Navarro
<> wrote:
> The dcumentation for (granted, unofficial, by the University of
> Missouri/Kolenikov, but on the search,net index) the package `annfit'
> is very minimal. Does anyone have more information on what this
> actually does? Is it a textbook directed-graph logistic-transformed
> backpropagation network? Is the `units()' option re: the number of
> units in the hidden layer?
> From its behaviour it's apparently not fully linear -- i.e. it applies
> a nonlinear transform after weighing at each unit -- but it could be
> anything from a cumulative normal to a hyperbolic tangent and still
> stay "textbook". (When coding ANNs by hand I've always preferred the
> hyperbolic tangent for anything that's not a binary classification
> problem, so it's hard to guess Kolenikov's preference regarding this.)
> As it is now it looks like a good way to cross-check dimensionality
> reduction procedures like (linear and nonlinear) PCA, but I can't
> quote these estimates if I don't know where they came from.
> *
> *   For searches and help try:
> *
> *
> *

Stas Kolenikov, also found at
Small print: I use this email account for mailing lists only.

*   For searches and help try:

© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index