it is time-consuming to Training deep neural network model on large dataset is time-consuming and abundant computing resources. We need to be clear that the purpose of cross validation is to get the optimal model on the current dataset. In addition, deep neural network has such a training trend that if the performance of the model is better in the early stage of training, the convergence performance of the model will be better. Therefore, when we train n models using different data split, we can select the N/3 models with better performance at 5th epoch, and the rest of training models stop. Then, we can choose the N/ 3 model with better performance at 10th epoch. In this way, the iterative operation can output the model with the better performance as cross validation model.

K-fold is the best approach as every sample of the data is validated but in case of deep learning i can recommend to use standard data splitting procedure where 60% , 20%, and 20%, can be used as training, testing, and validation set.

Thanks.

Dolly Das

Cross-validation is regulated when training the neural network.First, it is about 5% of the data, then with the quality control of training, increasing to 20%. It is no longer viable. Good training software packages use Genetic Algorithms to determine the ratio: "training" - "cross-checking"- "testing".The speed of learning is very dependent on the quality of the program used and the power of the computer. There are programs that use GPUs and / or multi-core processors to speed up calculations. Thus, the concept of "Deep Learning" is not a standard. For some conditions it is 10000 data, for others it is 10000000 data.

@Huang Bin Kindly reestablish the correlation of epochs with N/3. Sounds repetitive and confusing.