Why does neural network accuracy grow with error?

Question

Neural networks

I am solving an image classification problem. To debug the rest of the code, I wrote a simple convolutional neuron of 5 convolutional and 1 fc layers. On training, I gave out the following graphs: train and validation accuracy and error. Why, in general, does the error on the validation set increase with the accuracy on it?

There are 42 unique classes in total, ~20k pictures, 20% of them are the validation sample

PS Correct me if I'm wrong, but the fact that the validation error is growing indicates network retraining, right?

Charts:

Answer 1

Yes, it's definitely a retraining.
In order to answer why the accuracy also increases, you need a little more information about your project:

1. How much total training data?
   
2. How many unique classes?
   

Answer 2

try using dropout, it is quite possible that this is retraining