amit@csd4.csd.uwm.edu (amit srivastava) (09/27/90)
Dear netters:
I am new to neural networks and am trying to implement a neural network model
to solve the following. I would like to know if this is a reasonable task,
and I am seeking suggestions, comments, advice, etc.. I thank you in advance.
Pattern classification problem
------------------------------
I am using lumber as an example to demonstrate what I am trying to achieve.
There are a number of related problems in the real world to which this
approach can be taken. The goal here is:
- to solve the stated problem using neurocomputing.
- to gain an understanding of the neural network approach.
- to use the 'toy' problem to demonstrate the capabilities of this approach.
- to find out the limitations/problems with this approach.
- to vary different parameters in the model to see the model's
performance.
I would like to classify between types of lumber. Consider we have three
different types of lumber: birch, ash, and rosewood. Some of the features of
this lumber are weight, brightness, grain pattern prominence, straightness of
grain etc. These are the features. We can obtain numerical data for all the
features (Even something like grain pattern prominence can be quantified from
the magnitude and frequency of occurence of the light to dark transitions.)
The neural network model would be used to classify the different types of
wood.
Approach
--------
Several input patterns would be used to train the network. The learning phase
of the network will be achieved when the network has learned how to classify
the training samples. The network will then be fed with data, and we will
test the performance of the network in how well it classifies the wood.
Some more details
-----------------
Input vector -----> Output vector
x -----> System ------> 0 0 1 object is C
-----> ------> 0 1 0 object is B
-----> ------> 1 0 0 object is A
A B C
Possible configuration: (Will need to find out if this is feasible)
Input vector: Will use 2 to 4 features to classify the object (real numbers)
Output vector size : 2 to 3 (0 or 1)
Number of hidden layers : ???? (Maybe upto 2 - 3)
Model to be used: standard back propagation
Analysis
--------
Analyze the learning rate, tolerance to noise in input, accuracy of the model.
The network parameters that can be varied are number of hidden layers, number
of nodes per layer, learning rate.
Thank you all in advance.
amit@csd4.csd.uwm.edu
------------------------------------------------------------------------------
--
_/_
__. ____ o / |amit@csd4.csd.uwm.edu|
(_/|_/ / /_<_<__