Hyperparameter

Aspects of a neural network set by humans before training (e.g., number of layers/neurons, activation function) that remain constant during training.