Before we start to train our Network, and see if it can learn the XOR problem, we will need a couple of tools for visualizing the network's performance as it learns. To allow for this, we have created two objects called Logs, one for recording the state of various network statistics and variables as they are computed during the training process, and one for recording the activations of the hidden and output units, when the network is tested with the learning process switched off. The data recorded in these logs are displayed in the GraphlogView and the GridLogView respectively. Double click on both of these to open them up.
The GraphLogView is set to display the sum of the sum-squared-error
measure (sum_sum_se
), summing over all output units and all of
the training patterns presented within each epoch. The Epoch number
will be the X axis and the summed squared error will be the Y axis.
When we start to run the network this will be updated after every epoch
as we shall see.
The GridLogView is set to display detailed information about each pattern whenever a special test process is run. During the test, each event in the environment is presented once, and the View displays the epoch number, the Event name, the sum_squared error over all output units (though in this case there is only one), and the activations of the hidden and output units that occur when the input pattern associated with this event is presented.