Adaptive Networks Tutorial
Software by Vince Ramey - UC Berkeley, Nogales Lab
e-mail : vince_ramey at berkeley dot edu
phone : (510) 666-3335

-------------------------

Needed source code:
Edge.cpp
Edge.h
Node.cpp
Node.h
imagic.h
Adapt.cpp
makfile

-------------------------

Compiling:

Unpack the above files into a directory and use the 'make' command (linux and
mac) to build the binary.  This requires the g++ compiling tool, or you can
build the source by yourself however you want...

-------------------------

Using the program:

The default binary name is CAN (for Classification by Adaptive Networks).  Run
it with no arguments to get a usage message.  The parameters are as follows:

<input file> : name of the 'REAL', 4 byte imagic stack whose particles you wish
to classify

************NOTE: the input is expected WITHOUT suffix, so 'myImagicStack'
works but 'myImagicStack.img' doesn't*************************

<output name> : this will be the base filename for all output files (network
images, average images, class membership select files, and connection map file)

<iteration number> : Number of times data points will be presented to the
network.  ~15 times your particle number is sufficient.

<direct learning rate> : Fraction of the distance a network node will move
towards a presented data point.  0.01 works very well.  Higher values will make
the learning more noisy but possibly explore more of "image space".  Low values
may prevent the network from accurately capturing your data.  Feel free to
experiment!

<secondary learning rate> : Fraction of the distance a network node connected
to the highest scoring node will move toward that data point.  This effect
gives the network interesting "cohesive" properties and helps with
classification.  0.0005 is recommended.  Same goes for this value as the one
above.

<max age> : This is the number of iterations a connection between nodes will be
maintained while unused before it's abolished.  Higher values lead to a more
connected network, but over connection can lead to inhibition of correct
classification as the network will be over constrained.  Low values will make
an almost unconnected network, which can also hurt learning.  Values from 10 to
40 have been used.  I usually use 20-25.  Experiment here as well...

<number of nodes in the network> : Final number of points in the network.  This
also means that the data will be represented in this many "classes", although
because the averaging is different you may overshoot here and be fine.  Use a
similar number here as to what you would ask from an MSA classfication.

-------------------------

Output of the program:

(output_name).hed/.img - Final locations of each network point in image space
after classification.  Use these images as references for alignment as they
have the lowest noise

(output_name)_avg.hed/.img - After classification, each data point is assigned
to a network point based on distance.  These are then averaged and these images
are generated.  These are directly comparable to other methods since these are
actual averages of the data.

(output_name)_class***.spi - SPIDER select files for the averages generated
above.  This is here if you want to do something with individual classes, (use
them with tilted data for a Random Conical Tilt reconstruction or maybe edit
out bad particles IDed with classification method).

------------------------

Sample run of the program:

> CAN myImagicStack outStack 70000 .01 .005 20 200

This would classify the images in myImagicStack.hed/.img, putting the results
in outStack*.