Is CNN only applicable to time-series data or image data?
When should we use CNN instead of MLP?
$\begingroup$
$\endgroup$
1
-
$\begingroup$ A simple answer is when you need a convolution like applying filters in image processing or auto-regression in time series. $\endgroup$– OmGCommented Sep 17, 2021 at 22:23
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
CNN applies the same filters to every "chunk" of the input data. It's applicable when you think every chunk should be processed the same way.
For example, we think a face in the top-left of the image should be recognized just as well as a face in the bottom-right. So we do expect to process each part of the image the same way, and a CNN is good for this.
If we did not use a CNN, the network would need to learn separately at each position. It would learn to recognize faces in the top-left based on training examples where faces were in the top-left, and it would learn to recognize faces in the bottom-right based on training examples where faces were in the bottom-right. Because a CNN necessarily uses the same weights on each part of the image, seeing a face anywhere helps it update the shared weights which helps it recognize faces in every part of the image. Therefore, less training data may be needed.
A CNN might not be so good at classifying MNIST digits, for example, for the same reason. A /-shaped line in the top-left probably indicates a 6, but a /-shaped line in the bottom-right is more likely to indicate a 9, so we actually don't want the same weights there. It could still learn to sort this out in its final dense layer, of course. But if you wanted to find digits in a larger image you'd use a CNN.
