From my understanding of unsupervised DNNs for image classification:

  • The input layer is a 4,096 dimension vector (for 64 x 64 images)
  • The hidden layers represent much lower "features" as identified by the back propagation
  • As the model is generative, the output layer is also a 64 x 64 image

Therefore, how do can we make a prediction that a new unseen image contains a specific image class (e.g. cat) if we lack labelled data?


1 Answer 1


After a lot of reading, I think I now understand. We really need to build 2 models.

Model 1

  • Unsupervised
  • Lots of unlabelled images
  • Used to 'learn features' (i.e. better that we have done manually through years of research e.g. edge detection, colour features etc).

Model 2

  • Supervised
  • Few labelled images
  • Use model 1 as the 'feature extractor'. i.e. pass a training image through model 1 and use the output layer as the feature vector.
  • Use the same approach to test images e.g. model 1 to extract features, then use the second model to output label predictions

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.