Yes they play with various layers. Layers closer to the input act more like edge enhancers, while higher layers emphasize whole objects ("animal" enhancers). You get increasingly less syntactical and increasingly more semantic as you go deeper in the network.