The trained Lab 2 network looking at the scene. Each tile is one channel's ReLU response map, brightest where that filter's pattern is present; the grid is the layer's whole output volume, laid flat. Click a tile and it is laid over the photograph in gold. Work the layers in order: conv1's maps are edges and colours you could name from Chapter 1; by conv4 the maps are coarser, stranger, and increasingly about things — and no tile of conv4 can be explained by pointing at one filter, because its receptive field covers most of the frame.