C4W1L09 Pooling Layers

[@danielregassa9805]

For anyone wondering why average pooling isn't used often, it's because its functionality can be easily replicated by a filter with all elements = 1/n where n is the number of elements in the filter

[@ekayesorko]

thanks man.

[@WahranRai]

But backpropagation is in favor of pooling (no backpropagation)

[@ah-rdk]

good catch. also you can argue n=f*f.

[@jimmiemunyi]

4 years later. Thank youuuuuuuu

[@rajkkapadia]

Amazing explanation sir.

[@batoulZreik]

I have test and I am very grateful 🙏

[@safi2297]

it's really useful and easy to understand thanks for the video. keep it up the good work.

[@arkanandi8806]

Pooling layers incorporates to a certain extent spatial invariance. It would be really great if you can just describe why and how!

[@inquisitiverakib5844]

awesome content. I've a question if in the pooling layer no learning occurs then what is the need to do pooling

[@흑룡-d6n]

교수님. 컨볼루션 레이어 후에 값을 일자로 재배열후에 fc layer로 넣는건 알겠습니다. 그후 몇개의 층을거쳐 개수를 좀더 줄인후에 softmax하여 classify하는것도 알겠습니다. 그후 역전파는 어떻게 하는건지요? softmax하여 나온 값부터 어떤 기준을 가지고 역전파를 시작하여 fc layer를 거쳐 어떤식으로 convolution layer의 필터에 가중치를 적용하는지 그 과정이 생략되어있어 이해가 어렵습니다.

[@Vinoth89Karur]

Awesome sir.. Thank you so much..

[@ervinperetz5973]

Why do the number of channels double in AlexNet and VGG-19 ? Supposedly it's because overlapped pooling is used. But it's not clear how the extra channels are formulated. (e.g. for 2x2 overlapped pooling, presumably with stride 1 in both directions, width and height are halved (unlike in your overlapped pooling example), and #channels doubles; that doesn't add up wrt the number of pooling operations).

[@ati43888]

thanks

[@sandipansarkar9211]

nice explanation

[@davidtorres5012]

Thanks a lot !

[@juanandreslopezcubides5626]

If I have a dimension of 11x11 and a maxpool2d of 3, according to the formula it would be 9, but in Keras it says 3, why?

[@fariman-kashani]

Thanks! so useful.

[@nikilragav]

isn't average pooling actually a convolution -> box convolution

[@fatimahmath4819]

thank u very much sir

[@ahmeedhmdi7874]

[@MrQwerty2524]

So, does this formula mean that we substract 0.5 when dealing with decimals?

[@adityaachmad2265]

anyone know about backward pooling ?

[@codingtheworld674]

kzbin.info/www/bejne/jnaWnKWcaKiEotU After 2 Years :), but maybe someone also wants to look.

[@anandinamdar4054]

is maxpooling differentiable?

[@timharris72]

No there aren't any learnable parameters

[@NisseOhlsen]

yes, it is differentiable, but unless you made the size of the pooling window a parameter you wouldn't get anything out of differentiating. Example: If you have a 3 by 3 filter you have n = 9 parameters in your filter. so the max pooling operation would for once instance of a stride be f(x) = Sum(x)/n, where Sum(x) means the sum of all n elements. Differentiating with respect to x would give you Sum(1/n), which doesn't help you since x is not our parameter and this is not what we are trying to train. Differentiating with respect to n, allowing n to be trainable, would give you -n^-2. So IF we allow the pooling filter size to be trainable, the YES, max pooling is both differentiable AND usefully so. However, this is seemingly not used (although I'm not sure why).

[@shubhamchandra9258]

The entire neural network as a whole is differentiable. That can't happen if the pooling layer wasn't differentiable. Differentiable means for every small change in input there is a small change in output and not abrupt change.

[@FasstEddie]

is it just me or does the matrix glitch multiple times in this video?

[@sjl4554]

underlining reason of max pooling??

[@chaitanyag6297]

Max pooling, which is a form of down-sampling is used to identify the most important features. Means we are just taking the prominent edge/feature n that part , after conv layer edges will have a high positive number , so when you take the highest value in a part , you are looking at the edge/feature which is dominating others and more distinguishing , this has the advantage of downsizing our data for the dense layer to have less connections while taking the important features(leaving the less dominant features behind)

[@abishekseshan1757]

Chaitanya G But what is the guarantee that the pixel with highest value is most important? How can we determine that?

[@oscarw1976]

@@abishekseshan1757 max pooling isn't necessarily applied to pixels, it can be applied to a layer of neuron outputs

[@rohitborra2507]

@@chaitanyag6297 thanks bro

[@snippletrap]

@@abishekseshan1757 there is no guarantee but it works well in practice