for example one could compute the mean or max values a particular feature over region of damage and so you can use the summary statistics which are much lower dimension dimension compared to the extracted feature and day actually improve the the result for and learning because the now you have much less overfitting you have much less feature that you're learning and these aggregate operations are carpooling depending on the type the statistic its either mean pooling or max pooling and M one can choose the pooling region to be contiguous areas and image and only post feature generated from the same hidden you need and these unit will be translation invariant that's another property this mean that the same feature will be active even when the image undergo some small translation and this translation invariant features is a often desirable in many takes like objective reconition or even argue recognition alright to him what did I want to say I think now we have more less everything to describe the convolutional neural network which is new own network compo compose the one or more conclusion a layer often with some subs a sub something step and then followed by one or more fully connected layer and architecture offer convolutional neural network is designed to take advantage of that this to the structure of an input image and the 2d like an can also be like a speech signal for example and another benefit of the these conclusion your network is that they're easy to train and have much few were parameter and those fully connected to them newman et routes with the same number of hidden unit newman et route with the same number of hidden unit so by doing conclusions you end up with much your prime year so how do we do that in Python