while training a model no matter what kind of task it aims to achieve, we always need to split existing dataset to three parts, namely training set, validation set and test set.
Training set is used to train your model, that being said, used to calculate the gradient and update parameters; Validation set is used to evaluate the model performance. Now you might start wondering, these two sets sound perfectly enough to do the job of training the model, you can update the model parameter and you can evaluate the model performance, why bother using the test set?
Well, the aforementioned question makes perfect sense and I will explain the importance of test set in the following part.
To quote from Fchollet's new book Deep Learning with Python and I paraphrase, while developing a model, parameter tuning , such as hyperparameter selection, is always conducted using the feedback signal from the performance of the model on the validation data. To put it in a nutshell, some information is leaked from the valiadation set, and the model could quickly overfit on the validation set. This notion is called "information leak". To avoid information leak, we need a completely different and never-seen dataset to test your model's performance.