Convolutional neural network, why do we need multiple layers of reconciliation and how do kernels work on deep layers?

Question

Neural networks

I understood the principle of cores for the first reconciliation, this is to highlight some signs. After that, we get maps of a new image in which there will be sums of passage of nuclei through the image. Let's say I got n-th number of cards. In this case, how do new cores work for these cards, what is the point? After all, there are just numbers whose values ​​show the presence of a sign, but a set of these values ​​no longer gives some kind of sign, but only shows the presence of these signs. At this moment, I don’t understand, and even in articles and videos it is very vaguely shown, arguing that we are again highlighting signs. But in fact, I cannot understand the role of kernels on deep convolution layers.

Answer 1

If you understand ordinary multilayer neural networks, then it will be easiest to formulate that the "deep" convolutional layers play the same role as the "deep" layers in a conventional neural network. They take a set of features from the previous layer and combine them to get new higher-level features.
The difference is that the neuron of the typical Dense layer of the neural network owns all the information from the previous layer, and the convolutional layers operate on "local" kernels. Therefore, usually some kind of pooling is done between the convolutional layers, and then the kernels of the new convolutional layers, it turns out, use more information from the previous layers.

After all, there are just numbers whose values ​​show the presence of a sign, but a set of these values ​​no longer gives some kind of sign, but only shows the presence of these signs.

A set of two basic features can be considered as one high-level feature.
A very hypothetical example:

1. For example, on the first convolutional layer, you found the “left corner of the eye” with one core, and the “right corner of the eye” with the other core.
   
2. We did pooling. Now the new convolution layer with the same kernel size in pixels will cover most of the image
   
3. On the second convolutional layer, we combine the information that the "left corner of the eye" is next to the left of the "right corner of the eye", that is, on this layer we can say that this is a new feature of the "eyes".