Interpreting the graph (Overfitting / Loss / Accuracy)

[spanky...@gmail.com]

Hey Guys,

i want to discuss the interpretation of the shown graph. While "techniques to combat overfitting" already have been mentioned in [1] it is not my intention to discussed them here again. The only thing would like to know is: 

Why is the accuracy of the val set increasing while the loss of the same becomes more and more worse? 

Accuracy and loss of the training set show a strong correlation. If the loss becomes lower the accuracy is increasing. I would expect the same for the validation set.

Thanks in advance for every hint!

   

[1] https://groups.google.com/forum/#!searchin/caffe-users/Overfitting/caffe-users/ON93sHLbuJ0/bebGPPBHS3cJ

[spanky...@gmail.com]

By the way...the problem has 27 classes with approx. 1500 images per class.