Outliers are values or examples that are distant from most other observations in a dataset. They may result from errors, rare but valid cases, unusual inputs, or extreme model predictions, and they can distort training or evaluation.
Values distant from most other values. In machine learning, any of the following are outliers: - Input data whose values are more than roughly 3 standard deviations from the mean. - Weights with high absolute values. - Predicted values relatively far away from the actual values. For example, suppose that is a feature of a certain model. Assume that the mean is 7 Euros with a standard deviation of 1 Euro. Examples containing a of 12 Euros or 2 Euros would therefore be considered outliers because each of those prices is five standard deviations from the mean. Outliers are often caused by typos or other input mistakes. In other cases, outliers aren't mistakes; after all, values five standard deviations away from the mean are rare but hardly impossible. Outliers often cause problems in model training. Clipping is one way of managing outliers. See Working with numerical data in Machine Learning Crash Course for more information.