A nice little applet showing the perceptron learning rule in action. The points represent the training examples, with two possible target values (0 and 1). Each example has two input values (x_1 and x_2). If you tick the option "Vector", the applet shows you the vector of weights associated with the two inputs x_1 and x_2 (i.e. the vector consisting of the values w_1 and w_2). This vector, also called "normal" vector, is always orthogonal to the line that we try to learn to discriminate the two groups of points. There is also a bias input (always 1 in this applet). The bias weight associated with this constant input (i.e. the weight w_0) determines how far the line is from the origin. If w_0 is 0, then the line goes through the origin (i.e. the point where x_1 is 0 and x_2 is 0).
If the current example is mis-classified, the perceptron update rule updates the weight vector, which means the current x_1 and x_2 values are added to/subtracted from w_1 and w_2 (i.e. the example is added to/subtracted from the weight vector). This is shown in the applet. The applet uses a learning rate of 1.0, otherwise the update vector would be shorter or longer than the example vector.
http://isl.ira.uka.de/neuralNetCourse/2004/VL_11_5/Perceptron.html