Here is a quote from deeplearningbook which I am trying to process. I am not sure what do they mean by this quote, can someone help me understand please?

Pooling over spatial regions produces invariance to translation, but if we poolover the outputs of separately parametrized convolutions, the features can learnwhich transformations to become invariant to (see figure 9.9)


1 Answer 1


This will become clearer if you try to visualize the convolutional filter functioning.

Let's say you have a filter that is extracting horizontal lines from the image. When the filter extracts the dominance of horizontal lines in the image and assign a value in the generated matrix. Now we have the weights of the horizontal lines after the operation of the filter. Among all the horizontal weights, the most significant will be the one that has maximum weights so, we apply pooling (Max-pooling in this case) and discard all the other weights to select only that one having the maximum impact. By doing so, the filter selects only the features having maximum wights and discard the remaining. As the significance of the required object is from the line having maximum weight, the transformation is invariant. This also makes to tensor smaller in size and the process becomes faster too.

Perhaps, you can also refer the CS231 for a better understanding of the explaination.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.