In the control panel on the right hand side, you can enter, or select, learning rate, momentum, the number of training cycles, and the range of the weight initialization that is employed at the start of a simulation, and every time the ‘reset’ button is clicked.

The ‘Batch Update’ check box has the effect that weight changes are accumulated over the entire pattern set, and only made once, after all patterns have been seen.  If this is not checked, weight updates are made after every pattern presentation.  This default behavior has the grand title of ‘stochastic gradient descent’.

The ‘X-entropy’ checkbox affects the error calculation at the output units.  If checked, the cross-entropy error function is used.  This has a particularly simple derivative.  If not checked, the error at the output units is multiplied by the derivative of the logistic function.  Cross entropy is theoretically motivated when the outputs are to be interpreted as probabilities.

If there are no more than 24 individual patterns, they all appear in the combobox labelled ‘Pattern’.  This allows single patterns to be selected, and the ‘Test one’ to be clicked.  As can be seen below, the units display their individual activations, and the output is also provided in text in the small Console window at the bottom.  If there are more than 24 patterns, this combobox is unusable, and the name of the pattern file is displayed instead.

If ‘Test all’ is clicked instead, the average per-pattern error is computed over the whole training set and printed in the console.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s