Example 4: Perceptron Learning
in 2-d pattern space revisited
The following applet demonstrates
multi layer perecptron learning using the back-propagation
algorithm. Using this applet, you can train the perceptron
to learn in a 2-d pattern space. By placing three distinct
patterns of red dots and blue dots, the perceptron can learn to
distinguish between the patterns. An example is given below:
Instructions:
To Train The Perceptron:
Select the desired number of red and blue dots by
clicking the radio buttons provided.
Click on the white box on the left fo the applet to place
red dots. To place blue dots do the same but hold down
the <Shift> key.
Adjust the training parameters as desired. Legal values
are as follows:
Learning Rate: 0.0 to 1.0
Iterations: 1 to 10000
Error Threshold: 0.0 to 0.5
Click the Train button to begin a normal
training session OR
Click the Step button is not implemented
for this applet.
During Training:
To stop a training session (normal) in progress, click
the Stop button.
A few notes:
The progress of the training session is displayed
in the information field at the bottom of the
screen.
The error for the current input vector is
displayed in the Current-Error text
field.
x1,x2 are the 2 inputs of the multi layer
perceptron
w1,w2 are the weights effecting x1 going into the
2 hidden perceptrons
w3,w4 are the weights effecting x2 going into the
2 hidden perceptrons
The perceptron's ability to classify the inputs
into distinct classes (0(red) and 1(blue)) is
shown in a graph on the pattern space.
If either of the two training lines disappear and
dont return this IS NOT an error. It is due to
the algorithm finding a locoal minimum and not
being able to get out of it.
Some useful info:
To clear the Total error list and the pattern space click
Clear.
To plot a graph showing the results of training click Plot.
