September 2000

This directory contains the neural network classifier procedure nnet.pro and
training program train_nnet.pro and supporting procedures.    It is based on
the unpublished MS Thesis "Stellar Spectral Classification Through Neural
Networks" by T. Beck (ACC, Inc.)

nnet                 -neural network classifier
train_nnet           -neural network training program
nnet_write_weights   -Writes trained neural network weights to a FITS file
nnet_read_weights   - Reads trained neural network weights from a FITS file

Procedure:

1) Collect training data.   At least one good example of each type
2) Training data must be fit into a vector and normalized
3) Run train_nnet.pro.
4) Use trained weights to classify other data

_____________________
The example is taken from the prologue of train_nnet.pro

;*EXAMPLE:
;	This example uses the neural network as a stellar spectral classifier.
;	It could be used to classify any type of data, if the data could
;	be input as a normalized vector.
;	--------------------------------------------------------------------
;	You have a set of 10 flux & wavelength calibrated spectra. If 
;	necessary, resample the spectra to the same dispersion (eg. nm/pixel). 
;	Extract the same wavelength region from all spectra. Normalize. Make 
;	sure all pixel values are between 0 and 1.0. Stack all spectra into 
;	a single 2-D array. This is the training set (see input variable 
;	"train_set" above). If each spectrum has 200 pixels, then the size of
;	train_set will be (200,10). n_pat = 10 and n_in = 200 also.
;
;	Create a integer vector ("classes", above) of 10 elements, each 
;	element is a number that designates the spectral type of the 
;	corresponding spectra in the training set, by subscript:
;
;		classes(0) <====> train_set(*,0)
;
;	It is help to generate a lookup table:
;
;		class		SP type
;		-----		-------
;		  0		  M0V
;		  1		  M1V
;		  2		  M1.5V
;		  3		  M2V
;		  4		  M3V
;		  5		  M4V
;		  6		  M5V
;
;	Example of classes vector:
;
;		IDL> classes = [0,1,2,2,3,4,4,5,6,6]
;
;	Note that in this case some spectral types have more than one example.
;	It is a good idea to have a many examples of each spectral type as
;	possible, this will allow the neural net to generalize better and be
;	able to ignore noise. 
;
;	CAUTION: Two examples of the same spectal type that very different
;	in appearance due to noise, poor calibraion, etc. may cause the
;	network not to converge to a solution. 
;
;	In this example the number of output neurons (n_out) is equal to 7.
;	Set n_hid to some number between n_in and n_out, in this example,
;	100 would be a good choice.
;
;	Ready to run:
;		IDL> train_nnet, 10, 200, 100, 7, train_set, classes, $
;			bias_hid, w_hid, bias_out, w_out
;