Hiperparametre

Hyperparameter is the setting values that are automatically learned from the training data of a machine learning model, as opposed to the training process, determined by the man and shape how the model will learn. While the model parameters are learned from the data during training, hyperparameters determine the framework of this learning process: learning rate (learning rate, how much the model will change weight in every step), stack size (batch you), layer count and width are typical hyperparameter examples, such as regularization (regularization).

The correct selection of hyperparameters affects model performance dramatically: a very high learning rate can lead to the model to make unstable swings during training and no close to a good solution, a very low learning rate training can slow unnecessary location or cause the model to jam at a local minimum. In order to find these values, systematic methods such as grid search (grid search), random search (random search) or Bayesian optimization are used; this process is usually carried out together with cross verification, making sure the selected hyperparameters are overall good results.