# Back-propagation through max pooling layers

I have a small sub-question to this question.

I understand that when back-propagating through a max pooling layer the gradient is routed back in a way that the neuron in the previous layer which was selected as max gets all the gradient. What I'm not a 100% sure of is how the gradient in the next layer gets routed back to the pooling layer.

So the first question is if I have a pooling layer connected to a fully connected layer - like the image below.

When computing the gradient for the cyan "neuron" of the pooling layer do I sum all the gradients from the FC layer neurons? If this is correct then every "neuron" of the pooling layer has the same gradient?

For example if the first neuron of FC layer has a gradient of 2, second has a gradient of 3, and third a gradient of 6. What are the gradients of the blue and purple "neurons" in the pooling layer and why?

And the second question is when the pooling layer is connected to another convolution layer. How do I compute the gradient then? See the example below.

For the topmost rightmost "neuron" of the pooling layer (the outlined green one) I just take the gradient of the purple neuron in the next conv layer and route it back, right?

How about the filled green one? I need to multiply together the first column of neurons in the next layer because of the chain rule? Or do I need to add them?

Please do not post a bunch of equations and tell me that my answer is right in there because I've been trying to wrap my head around equations and I still don't understand it perfectly that's why I'm asking this question in a simple way.

• With regards to your first question, backpropagation is to see what weights and inputs influences your loss in what way. In case of max pooling only the max of the neurons influences the output (except for when there is a tie). So only propagate the error to the neuron that had the maximum activation value. Aug 21, 2016 at 10:53
• Yes, I understand this and I also said this in a recap at the beginning of my post. But I don't understand how do I "combine" the gradient of the next layer neurons to propagate back. Hope you know what I mean. Aug 21, 2016 at 13:03

If this is correct then every "neuron" of the pooling layer has the same gradient?

No. It depends on the weights and activation function. And most typically the weights are different from the first neuron of the pooling layer to the FC layer as from the second layer of the pooling layer to the FC layer.

So typically you will have a situation like:

$FC_i = f(\sum_j W_{ij} P_j)$

Where $FC_i$ is the ith neuron in the fully connected layer, $P_j$ is the jth neuron in the pooling layer and $f$ is the activation function and $W$ the weigths.

This means that the gradient with respect to P_j is

$grad(P_j) = \sum_i grad(FC_i) f^\prime W_{ij}$.

Which is different for j=0 or j=1 because the W is different.

And the second question is when the pooling layer is connected to another convolution layer. How do I compute the gradient then?

It does not make a difference what type of layer it is connected to. It is the same equation all the time. Sum of all the gradients on the next layer multiplied by how the output of those neurons is affected by the neuron on the previous layer. The difference between FC and Convolution is that in FC all neurons in the next layer will provide a contribution (even if maybe small) but in Convolution most neurons in the next layer are not at all affected by the neuron in the previous layer so their contribution is exactly zero.

For the topmost rightmost "neuron" of the pooling layer (the outlined green one) I just take the gradient of the purple neuron in the next conv layer and route it back, right?

Right. Plus also the gradient of any other neurons on that convolution layer which take as input the topmost rightmost neuron of the pooling layer.

How about the filled green one? I need to multiply together the first column of neurons in the next layer because of the chain rule? Or do I need to add them?

Add them. Because of the chain rule.

Max Pooling Up to the this point, the fact that it was max pool was totally irrelevant as you can see. Max pooling is just the that the activation function on that layer is $max$. So this means that the gradients for the previous layers $grad(PR_j)$ are:

$grad(PR_j) = \sum_i grad(P_i) f^\prime W_{ij}$.

But now $f = id$ for the max neuron and $f = 0$ for all other neurons, so $f^\prime = 1$ for the max neuron in the previous layer and $f^\prime = 0$ for all other neurons. So:

$grad(PR_{max neuron}) = \sum_i grad(P_i) W_{i\ {max\ neuron}}$,

$grad(PR_{others}) = 0.$