What is the advantage of using multiples of the same filter in convolutional networks in deep learning?

Question

What is the advantage of using multiples of the same filter in convolutional networks for in-depth study?

For example: We use 6 filters of size [5.5] at the first level to scan image data, which is a size matrix [28.28]. The question is why we do not use only one size filter [5.5], but use 6 or more of them. As a result, they will scan the same pixels. I can see that the random weight may be different, but the DL model will adapt to it anyway.

So, what is the main advantage and purpose of using multiple filters of the same form, and then in convnets?

Answer 1

Why is the shape of the filter the same?

Firstly, the shape of the kernel is the same, just to speed up the calculations. This allows convolution to be applied in a package, for example, using col2im transform and matrix multiplication. It is also convenient to store all weights in a single multidimensional array. Although mathematically you can imagine using several filters of various shapes.

Some architectures, such as the initial network, use this idea and apply different convolutional layers (with different cores) in parallel and add function maps at the end. This has proven to be very helpful.

Why is one filter not enough?

, , , , . . , .

, , , ... , . , . , MNIST.

?

: . , . , .

: , . , . , , , , . -.

, - , : . , , . , .

CS231n , .

Answer 2

Zeiler Fergus https://arxiv.org/pdf/1311.2901.pdf , .

​​ , () . 6 "" . , , , (, , ), . . .

Prewitt https://en.m.wikipedia.org/wiki/Prewitt_operator ​​3x3, , ( ) . , . , .