In CNN we have convolution layer and pooling layer for feature extraction. How the features are extracted in Fully connected neural network? Secondly CNN has more expressive power and and number of learn able parameters are also less in this case. I want to know that if CNN are so perfect then why we use fully connected neural networks?
2
-
$\begingroup$ Convolutional layers are useful if the input is either image and/or temporal based, e.g. data adjacent to each other are likely "related". Fully connected layer does not have that kind of assumption. $\endgroup$– pangyutengCommented Nov 23, 2018 at 13:54
-
$\begingroup$ If you find an answer somewhere, please post it here. $\endgroup$– Cloud ChoCommented Feb 2, 2023 at 23:48
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
2
In the case of classification, neural networks learn a coordinate system for the target classes which is a nonlinear mapping of the inputs. See: Can't deep learning models now be said to be interpretable? Are nodes features? These are "features" in the sense that they are an abstraction from the input, but they’re not really interpretable as some latent phenomenon about the input.
CNNs aren't perfect. For example, small modifications to the image can dramatically shift the classifier's disposition. But CNNs are only useful when the input has some sort of structure to it; for an image, that structure is that nearby pixels comprise an important shape, texture or edge. For a time-series, the structure is that yesterday is more similar to tomorrow than last year. FCNs are useful when the data is "tabular" in format, recording different measurements for an observation which are not related (spatially, temporally, or otherwise).
answered Nov 23, 2018 at 13:52
-
$\begingroup$ what do you mean by CNN is useful when nearby pixels are probably representing the same object?Can you please explain it a bit more? And can we say that we don't learn features in Multi layer network? $\endgroup$– hudaCommented Nov 23, 2018 at 14:01
-
$\begingroup$ In a natural image, nearby pixels are correlated. CNNs are good at learning filters to detect edges, textures and other structure. These are aggregated up across different image patches. The link in the post develops interpretation of FCN features in more detail. $\endgroup$– Sycorax ♦Commented Nov 23, 2018 at 14:09