A neural network model's training takes days; to speed up training what should you do?

Subsampling the training dataset reduces the amount of data the model processes during each epoch, significantly decreasing training time. While this might lead to a slight reduction in model performance if the subsample isn't representative, it's a common and effective strategy for speeding up initial training or experimentation with large datasets. Subsampling the test dataset is incorrect because the test set is used for evaluation, not training, and reducing its size won't speed up training. Increasing the number of input features or increasing the number of layers in your neural network would both increase model complexity and computational requirements, thereby increasing training time, not decreasing it.