Generalization of neural networks

Generalization is aimed at fitting the training data. It is an extension of the training we have done on the neural networks model. It seeks to minimize the sum of squared errors of the model on the training data (such as using ordinary least squares) and reduce the complexity of the model.

The methods of generalization are listed here:

• Early stopping of training
• Retraining neural networks with different training data
  • Using random sampling, stratified sampling, or any good mix of target data
• Training multiple neural networks and averaging out their outputs