I guess the common rule of thumb when choosing between CNN vs. DNN is if it has to do with images choose CNN and data points for DNN.

But what if input images are pretty small.. in my case (33,45,3) or (33,15,3), or flatten to 4455 or 1485 datapoints. .

When considering each as pixel as a datapoint I don't see that big of an issue using a DNN to train the data, given the number of data points are pretty small...

Why shouldn't I do this?

CNN: Convolution neural network

DNN: Deep neural network

  • $\begingroup$ Welcome to the site. Please spell out CNN and DNN $\endgroup$ – Peter Flom May 21 '17 at 12:19
  • $\begingroup$ ?.cnn = Convolutional neural network, DNN = Deep neural network?.. Something i've misunderstood here? $\endgroup$ – Vej May 21 '17 at 12:40
  • 1
    $\begingroup$ Convolutional networks can be deep. Do you know what makes a conventional network convolutional? $\endgroup$ – generic_user May 21 '17 at 12:47
  • $\begingroup$ @generic_user yes the convololutional layer.. $\endgroup$ – Vej May 21 '17 at 13:18

It depends on your goal and how the data looks like. If, for example, your goal is to detect some object in the image, regardless of its position, then you would use a CNN, because the convolutional filters, together with pooling, achieve that, and generalize much better. All successful models which have won some competition (imagenet, R-CNN, ResNet, ...) are based on convolutional layers, among other things.

By not encoding invariances in the architecture, you need to collect much more data, use more parameters (you need to be able to detect the very same objects at several positions by different groups of weights), and most likely it will not perform as well.

| cite | improve this answer | |
  • $\begingroup$ but what if the number of pixels is so small.. like here where I have under 5000 data points.. Clearly a DNN would be able to handle that? $\endgroup$ – Vej May 21 '17 at 13:05
  • $\begingroup$ Basically I agree with the answer. @Vej, please take into account that a typical toy example used in courses for CNN is the cifar10, images of 32x32, and the CNN works pretty well compared to a Multi Layer Perceptron (I guess this is what you actually mean by DNN). Cheers. $\endgroup$ – lrnzcig May 22 '17 at 14:19

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.