My validation loss is bumpy in CNN with higher accuracy. Does my model overfitting?

Question

My training loss is constantly going lower but when my test accuracy becomes more than 95% it goes lower and higher. Does this mean that my model is overfitting or it's normal? What should I do?

My network has around 70 million parameters. And batch size is 16.

Here is my test and validation losses. What should I do? Also my validation loss is lower than training loss? Why is that?

Test training loss

I've used different kernel sizes and tried to run in lower epochs. Maybe I should train the network with more epochs?

Answer 1

Let's answer your questions in order. First about "accuracy goes lower and higher". This means that you have reached the extremum point while training the model. That is, your model has learned. Since your metric shows quite high indicators on the validation set, so we can say that the model has learned well (of course, if the metric is chosen correctly for the task).

Now about "my validation loss is lower than training loss". It's okay due to in essence of validation. I recommend you study what a validation, training and test set is.