Preprocessors normally have default range limits from 0 to 1. If you don't apply normalization to a dataset with a wide range of numeric values (when feature values that are too low and too high are present), then the neural network will tend to favor the feature values that have high numeric values. Hence, the accuracy of the neural network could be significantly reduced.

If values are spread across symmetric intervals such as (0,1), then all the feature values are considered equivalent during training. Hence, it also has an impact on the neural network's generalization.

The following are the preprocessors that are provided by ND4J: 

• NormalizerStandardize: A preprocessor for datasets that normalizes feature values so that they have a zero mean and a standard deviation of 1.
• MultiNormalizerStandardize: A preprocessor for multi-datasets that normalizes feature values so that they have a zero mean and a standard deviation of 1.
• NormalizerMinMaxScaler: A preprocessor for datasets that normalizes feature values so that they lie between a minimum and maximum value that's been specified. The default range is 0 to 1.
• MultiNormalizerMinMaxScaler: A preprocessor for multi-datasets that normalizes feature values that lie between a minimum and maximum value that's been specified. The default range is 0 to 1.
• ImagePreProcessingScaler: A preprocessor for images with minimum and maximum scaling. The default ranges are (miRange, maxRange) – (0,1).
• VGG16ImagePreProcessor: A preprocessor specifically for the VGG16 network architecture. It computes the mean RGB value and subtracts it from each pixel on the training set.