How many data do you need for training an Artificial Neural Network?

How many data do you need for the training of an Artificial Neural Network?

As much as necessary, but not more!

The number of necessary data for the representation of a problem is dependent on the complexity of the supposed functionality and the given problem but not from the type of the used model.

The confirmation on the sufficient number of data can be deduced by the accompanying validation procedures. In the case of missing data the according methods also support in the proper extension and generation of further data. That way Machine Learning methods can also be used for an effective design of experiments.

Approaching a problem that data centric way is much more natural and efficient in comparison to solution centric approaches like classical factorial design, where the number of data depends on the number of investigated parameters and the type and size of the model. Methods which claim to need less data for training/calibration of the models are often based on some analytical (pre-) assumptions. Here one must not forget to take into account the necessary data for the validation of the assumption as well. These data centric methods and approaches help to estimate the need of further training data and the density of the sufficient data for a given problem dependent on its complexity level.

For more detailed information about the given question please consult info@andata.at.

Last update on 2017-05-01 by Andreas Kuhn.

Go back