1
$\begingroup$

I am trying to build an image classification using transfer learning of VGG16 model. I acquired very small data set of 200 images for each class and used 10 images as validation(I know the data set is small but the requirement was for controlled environment). The model accuracy on unseen data is not great but the issue is with the probability it returns. The model predicts class with abnormally high accuracy even for wrong prediction ex(1.0,0.,2.8 e15) this is predicted softmax output for 3 classes. Apart from improving accuracy I want the probability to be distributed normally because two object from different classes might be present in the image where model should note the presence of object from different class and give less confident prediction. My question is what is wrong with the model, is it low number of data-set or variations of images in a class(I have included number of different smaller category under single higher class as per requirement example: vehicle contains images of car bike truck) . I am exploring data augmentation to increase number of images but doubtful if that statifies the requirement and solving probability issue. Please help.

$\endgroup$
  • $\begingroup$ In other words, your pre-trained model fails to converge (reach small training error) on the new, small dataset? $\endgroup$ – Jan Kukacka Mar 14 '18 at 13:00
  • $\begingroup$ With new distinct image the model favors one particular class and probability of other classes are equal to zero. Even when correctly classifying the probability distribution is similar. $\endgroup$ – Unbanned Mar 14 '18 at 13:08
  • $\begingroup$ So the training images are classified correctly but unseen test images are misclassified? $\endgroup$ – Jan Kukacka Mar 14 '18 at 13:11
  • $\begingroup$ So there isn't really any over fitting, and you are unhappy that it provides "confident" correct predictions on unseen data? $\endgroup$ – Jan Kukacka Mar 14 '18 at 13:19
  • $\begingroup$ Side note: softmax outputs have nothing to with the prediction confidence, search this site for some relevant QA about this. $\endgroup$ – Jan Kukacka Mar 14 '18 at 13:22

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.