From the course: Deep Learning: Getting Started

Validation and testing

From the course: Deep Learning: Getting Started

Validation and testing

“

- [Narrator] As we build models, we need to also validate and test them against independent data sets to measure out of sample error. During the input preparation process, we usually isolate validation and test data sets for this purpose. What is validation? While performing, learning, and model improvement, we are comparing the predictions provided by the neural network for the training samples against its actual values and measuring errors. However, this is an in sample error and the model has no guarantee that it will perform the same against independent datasets. So after each epoch is completed and the weights and biases updated, we will also use the network to predict for the validation data set. We will measure accuracy and loss for the validation data set, and also investigate the same to make sure that it does not deviate significantly from the in sample errors observed. The model can be fine tuned, and the learning process repeated based on the results seen against the validation dataset. The final step in model building is evaluation. After all the fine tuning is completed, and the final model obtained, the test data set is used to evaluate the model. This is done only once at the end. The evaluation results are then used to measure the performance of the model in terms of its final accuracy and error rates.

Contents