In machine learning, there is two different conception to assess your model——validation data and test data.
Let's suppose you are a research on machine learning, you have a lot of data with label, maybe an image. You want to choose a good model for the task of understanding the image. You have model A, B, and C.
You separate your data into tree part.P1,P2 and P3. P1 for training. Now you got three network with weight.
You need to decide which one is best and fit for real world data. You run your network with data P2, then you have the lost1, lost2, and lost3. Suppose lost1 is the smallest. Then you run your network with data P3, and lost4.
You may write a persuasive paper now.
validation data——choose your model/fine tune your hiper-parameters
test data——assess your model
why two of them?
Because choose your model by test data can only prove it is best for these crop of data, or these batch of data. You have to choose other data to prove your model is good too.So the one-test lost will lower than validation-test lost.