I know how pooling works, and what effect it has on the input dimensions - but I'm not sure why it's done in the first place. It'd be great if someone could provide some intuition behind it - while explaining the following excerpt from a blog:

A problem with the output feature maps is that they are sensitive to the location of the features in the input. One approach to address this sensitivity is to down sample the feature maps. This has the effect of making the resulting down sampled feature maps more robust to changes in the position of the feature in the image, referred to by the technical phrase “local translation invariance.”

What's local translation invariance here?


Pooling has multiple benefits

  • Robust feature detection.
  • Makes it computationally feasible to have deeper CNNs

Robust Feature Detection

Think of max-pooling (most popular) for understanding this. Consider a 2*2 box/unit in one layer which is mapped to only 1 box/unit in the next layer (Basically pooling). Let's say the feature map (kernel) detects a petal of a flower. Then qualifying a petal if any of the 4 units of the previous layer is fired makes the detection robust to noise. There is no strict requirement that all 4 units should be fired to detect a petal. Thus, the next layer (after pooling) captures the features with noise invariance. We can also say it is local translation invariance (in a close spatial sense) as a shifted feature will also be captured. But also remember translation invariance in general is captured by the convolution with kernels in the first place. (See how 1 kernel is convolved with the whole image)

Computational advantage

The dimensions of the inputs in image classification are so huge that the number of the multiplication operation is in billions even with very few layers. Pooling the output layer reduces the input dimension for the next layer thus saving computation. But also now one can aim for really deep networks(number of layers) with the same complexity as before.


In addition in general it somewhat aides in detection as only the strongest feature feature filter is activated so in a sense it removes additional information.

But it obviously has draw backs resulting in combinations of features being detected which aren't actual.objects.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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