Patterns Processed per Weight Update Cycle

The number of training patterns to process per weight update cycle can have a large effect on the overall training process and convergence behavior. Qnet’s default is to process all training patterns prior to updating network weights (value is set to 0). Qnet computes a global error vector to be applied to the weight adjustment algorithm. This practice generally leads to the most orderly decent of both training and test set errors at the price of slightly slower training performance. This default method is strongly recommended for training sets where imprecise or fuzzy relationships exist between the inputs and outputs.

When weights are updated after a partial set of patterns have been processed, several distinct differences may be noted during the training process. Learn rates can play a larger factor on the observed level of generalized vs. memorization learning. Lower learning rates can improve generalization. It is also advised that Learn Rate Control (LRC) be turned off. The advantage of this method of network training is that weight updates are performed more frequently and network convergence times can be improved for some models.

The number of patterns processed per weight update cycle is set in the Training Setup/Training Parameters. A value of 0 (the default) processes all training patterns per update cycle. A value of 1 processes one pattern for each update cycle, etc. This value may also be altered during training. IMPORTANT: Changing this parameter during training can dramatically alter learning characteristics. It strongly recommended that the network be saved (or AutoSave be active) prior to changing.