You might be familiar with the Max Pooling which takes maximum values to down sample the feature map; lesser parameters for being robust to changes.
However, maxpooling can be replaced with convolutional layers. Why? cause some feature information can be lost when running through maxpooling layers. The medium article shows an example of image reconstruction using VAE to compare models with a maxpooling layer and a convolutional layer. As you can see in that article, the convolutional pooling outperforms the one with the maxpooling.
Addition to this, STRIVING FOR SIMPLICITY: THE ALL CONVOLUTIONAL NET says that
We find that max-pooling can simply be replaced by a convolutional layer with increased stride without loss in accuracy on several image recognition benchmarks.
when pooling is replaced by an additional convolution layer with stride r = 2 performance stabilizes and even improves on the base model
Reference
-
https://medium.com/@duanenielsen/deep-learning-cage-match-max-pooling-vs-convolutions-e42581387cb9
-
https://github.com/DuaneNielsen/maxpoolvsconv
- https://arxiv.org/pdf/1412.6806.pdf
- https://stats.stackexchange.com/questions/387482/pooling-vs-stride-for-downsampling