0
$\begingroup$

Can CNNs be used with input data which is not an image? The reason I'm asking is because the original image is often clipped in size because of border effects when doing the convolution.

But if the input is not an image, and I really need to use all the input data, is it possible to overcome the problem?

Improve this question
$\endgroup$
3
  • $\begingroup$ To what "problem" are you referring? $\endgroup$
    – whuber ♦
    Oct 23, 2014 at 3:22
  • $\begingroup$ typical problem of classification in ML. not necessarily applied to images or video. $\endgroup$
    – Bob
    Oct 23, 2014 at 6:04
  • $\begingroup$ CNNs exploit topological structure in the input variables. Do your input variables have topological structure? $\endgroup$
    – generic_user
    Jan 30, 2018 at 19:45

2 Answers 2

Reset to default
5
$\begingroup$

Convolutional networks work so well because they exploit an assumption about with weight sharing. This is why they only work with data where that assumption hold.

The assumption is a spatial one. It is best explained with a picture, where you do not care where exactly something is, which is sometimes called translational invariance.

As long as that assumption holds on your data, you can apply it. Other modalities are e.g. audio or (to some extent) text.

Improve this answer
$\endgroup$
2
  • $\begingroup$ I'm not sure if I understand. In all tutorials I've seen, the input is always cropped at the boundaries. Are you saying that is not the case? $\endgroup$
    – Bob
    Oct 25, 2014 at 0:01
  • 1
    $\begingroup$ I don't say that. For convnets, you need to bring all inputs to the same spatial dimension, though. $\endgroup$
    – bayerj
    Oct 25, 2014 at 17:58
2
$\begingroup$

Nothing in the CNN method requires clipping of the input data - that is simply a design choice by the modeller.

I am not completely sure what you are asking for, but it seems that you are interested in the problem associated with the edges of the input matrix. Each pixel at the edge of the image is only captured in a single neuron in the first convolutional layer, making the information less used for classification (which reduces accuracy). Zero-padding is often used to overcome this problem, and could easily be used for other data types as well.

Improve this answer
$\endgroup$

Your Answer

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

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