Problem

Features whose values exist over a wide scale negatively impact the loss reduction process of neural networks, resulting in longer training times and excessive use of processing resources.

Solution

All numerical features in a dataset are brought within the same scale such that the values are either between 0 and 1 or between -1 and +1.

Application

Statistical techniques, such as min-max scaling, mean normalization and z-score standardization, are applied that convert the feature values in such a way that the values always exist within a known set of upper and lower bounds.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

A data scientist prepares a training dataset that comprises features with varying scales. Feature A’s values are between 0 and 5000, Feature B’s values are between -15 and +15, and Feature C’s values are between 30 and 60 (1). The dataset is exposed to the min-max scaling technique (2). The resulting scaled dataset contains values such that all feature values are now between 0 and 1 (3). The scaled dataset is then used to train a neural network (4, 5). The network training process is completed in a shorter time period (6).