I have a dataset of 34k (200x200) images and I want to build an 8 class age detector.
I've tried a lot of different networks design, regularizations, dropout layers, grayscale images, data augmentation on the training set, learning rate, shuffle the data order etc... but I only get accuracy from 43% to 48%. I've also tried some famous networks as AlexNet or LeNet-5 but the maximum accuracy I can get is about 49%.

In such a task I would expect an accuracy similar to what a human can get or at least something about 80-85%. What can be the reason for such a low accuracy? What could I try?

  • $\begingroup$ The first thing I would suggest to look at is the confusion matrix, to analyze what kind of error happens. I don't think accuracy is a good measure for this, the classes are certainly imbalanced. Also are you sure that a human would reach 80-85%? $\endgroup$
    – Erwan
    Sep 16 at 10:15


Your Answer

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