why do we need to flip kernel before using conv2 in CNN?

We know that function conv2 can prefom convolution (between image and kernel ) and flip kernel before apply convolution to image according to defnition of convolution

y = conv2(image, kernel, 'valid')

.However, in convolution neural network(CNN) ,they flip the kernel before the use conv2

kernel = rot90(kernel, 2);

y = conv2(image, kernel, 'valid');

which means the kernel flip twice and this correlation not convolution why

[Jonas]

not that 90° is not flipped, flipping would be rotation by 180 degress or using flipud(fliplr())

and where exactly did you find this code (file, line number)

[Mohammedee]

function y = Conv(x, W) % % [wrow, wcol, numFilters] = size(W); [xrow, xcol, ~ ] = size(x); yrow = xrow - wrow + 1; ycol = xcol - wcol + 1; y = zeros(yrow, ycol, numFilters); for k = 1:numFilters filter = W(:, :, k); filter = rot90(squeeze(filter), 2); y(:, :, k) = conv2(x, filter, 'valid'); end end

[Mohammedee]

Look to this code..kernel rotated 180 Then pass it to conv2 And we know that conv2 will rotate kernel 180 again... This mean kernel rotated twice 180..

Answer 1

[Matt J]

The field of neural networks uses the term "convolution" loosely. There are other differences as well. We also know that in traditional DSP theory, convolution operations don't contain a stride parameter, but in the NN world, they do.

[Mohammedee]

It is not clear..

[Matt J]

Basically, neural networks researchers are not using the terminology "convolution" in the classical way. A true convolution in the original sense of the word should include a flip and should never have stride>1. That's the way convolution was originally defined. Without the flip, it should be called correlation, as you say.

[Mohammedee]

http://ufldl.stanford.edu/tutorial/supervised/ExerciseConvolutionAndPooling/

In this link i find other related answer but also not clear they said:

If you use conv2(image, W), MATLAB will first "flip" W, reversing its rows and columns, before convolving W with image, as below:

⎛⎝⎜147258369⎞⎠⎟−→−flip⎛⎝⎜963852741⎞⎠⎟(123456789)→flip(987654321)

If the original layout of W was correct, after flipping, it would be incorrect. For the layout to be correct after flipping, you will have to flip W before passing it into conv2, so that after MATLAB flips W in conv2, the layout will be correct. For conv2, this means reversing the rows and columns, which can be done by rotating W 90 degrees twice with rot90 as shown below:

[Mohammedee]

The answer here but i do understand that W will be not corrected after flipping.. Thus u have to flip before use. Conv2

(If the original layout of W was correct, after flipping, it would be incorrect. For the layout to be correct after flipping, you will have to flip W before passing it into conv2, so that after MATLAB flips W in conv2, the layout will be correct)

[Matt J]

If you use conv2(image, W), MATLAB will first "flip" W, reversing its rows and columns

Yes, conv2 will flip W internally and that is the correct thing for it to do, because that is the way convolution is defined. This definition ensures that conv2(1,W) = W. Example:

W=[1 2;3 4]

W = 2×2

1 2 3 4

conv2(1,W)

ans = 2×2

1 2 3 4

If you were to flip W manually, prior to giving it to conv2, it would mess this up:

conv2(1,rot90(W,2))

ans = 2×2

4 3 2 1