Keras: Unlogical validation lost values

Question

The following net produces unlogically validation loss values. For the training_loss it holds 0 < training_loss < 1 and for the validation_loss > 30.