1
$\begingroup$

Let's say I want to classify a dataset of handwritten digits (CNNs on their own can get 99.7% on the MNIST dataset but let's pretend they can only get 90% for the sake of this question).

Now, I already have some classical computer vision techniques which might be able to give me a clue. For instance, I can count the intersection points of the pen stroke

  • 1,2,3,5,7 will usually have no intersection points
  • 6,9 will usually have one intersection point each
  • 4,8 will usually have two intersection points each (usually a 4-way crossover yields two intersection points which are close together)

So if I generate some meta-data telling me how many intersection points each sample has, how can I feed that into the CNN training so that it can take advantage of that knowledge?

My best guess is to slot it into the last fully connected layer just before classification.

Improve this question
$\endgroup$
4
  • 1
    $\begingroup$ You could add the meta-data along all the pixels to your feature vector and let the model learn the correlation between number of intersections and digits. $\endgroup$
    – Feodoran
    Feb 18 '20 at 13:06
  • $\begingroup$ @Feodoran thanks. Is that different from what I said about adding it into the last fully connected layer? Is "feature vector" different than that? Just makings sure I've got the terms right. $\endgroup$
    – Alexander Soare
    Feb 18 '20 at 13:13
  • 1
    $\begingroup$ It is different in that I suggest a single, combined feature vector feeding into the first layer, while you are suggesting two feature vectors feeding into two different layers of the model. You have to try which one works better. $\endgroup$
    – Feodoran
    Feb 18 '20 at 13:20
  • $\begingroup$ Great. Thanks for that. $\endgroup$
    – Alexander Soare
    Feb 18 '20 at 13:22
1
$\begingroup$

"Distilling the Knowledge in a Neural Network" by G. Hinton et al. https://arxiv.org/abs/1503.02531

Improve this answer
$\endgroup$
1
  • 2
    $\begingroup$ Thanks for your answer -- it would be great if you could add a few more details about that paper apart from just the title and a link! $\endgroup$
    – Oliver Mason
    Feb 18 '20 at 11:06

Your Answer

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

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