I have 2 million samples and using a 60/20/20 split for train, validation and test.
Validation and train accuracy are very close to each other at 80% and 85% respectively but test accuracy much lower at 55%.
I know that we aren't supposed to monitor the test losses for each epoch but I checked and the test loss was increasing after 10 epochs but train and validation losses were both decreasing.
What am I doing wrong?
