Understanding the layers in CNN

Question

I am new to NNs and I have a question about the convolutional layers in CNN. A convolution layer is said to perform feature extraction or work as a feature extractor in CNN. The first convolutional layer learns to extract low-level features which means that the first layer will convert the original image into multiple copies of the original image (depending on the number of filters used) that contain only the low-level features, neglecting any other features in the original image, right ?

How does the next layers extract other features high level features i.e. faces if they were already neglected by the first layer ?

Why don't we have several separate layers, each of which does a specific extraction of some features and once all the needed has been extracted, the output is merged before the fully connected layer ?

I don't get the point of applying different filters (after the first layer)on the entry already filtered.

Answer 1

This is a good question to ask as it pertains to important neural network concepts.

"Why don't we have several separate layers, each of which does a specific extraction of some features and once all the needed has been extracted"

The answer is two-fold.

First, the layers depend on each other. The reason deeper layers can extract higher-level features is that they build upon the lower features extracted in the earlier layers. The layers may be different in architecture but they could simply be the same layer duplicated multiple times and initialized with the same weights.

Second, it is important to understand that the layers are not "told" what type of features to extract. Realizing that each layers represented various level of features was very important in better understanding the behaviour of neural networks and to confirm that no, it isn't all black magic. And it is important to understand that is a learning artefact which may not always occur depending on the network, data, and training strategy involved.