The https://arxiv.org/pdf/1412.6806.pdf Sriving for simplicity talks of how we could replace max pooling with an additional convolutional layer and stride r = 2. But can the same be achieved with average pooling instead ?
Asked
Active
Viewed 894 times
1 Answers
0
The point of the article is to show that the pooling operation can be learned instead of being hardcoded:
We emphasize that that this replacement can also be seen as learning the pooling operation rather than fixing it
Even though they only experimented with max-pooling the conclusion seems transposable to other pooling like average pooling as the replacement convolutional layer will likely learn this transformation.
The point is that a simple convolution with stride 2 can replace any pooling operation.
Louis Lac
- 5,298
- 1
- 21
- 36