Chapter 4 · The deep learning era (2006–2016)
Convolutional neural networks
CNNs stack convolutions so early layers see edges and later layers see eyes, wheels and faces.
Chapter 4 · The deep learning era (2006–2016)
CNNs stack convolutions so early layers see edges and later layers see eyes, wheels and faces.
A doodle goes in
Each layer looks at the layer before it. The first sees tiny lines. The next combines lines into shapes. The last names the whole thing.
One filter finds one pattern. A convolutional layer runs many filters side by side, each making its own feature map: one for vertical lines, one for diagonals, and so on.
Draw on the grid or pick a preset. Watch layer 1 light up where each kind of line is, then see which shape detector wins.
Output: this looks like “+”.
A teaching version: the filters and the shape rules are set by hand to know only four shapes. A real CNN learns its filters from many labelled pictures.
What do the first layers of a CNN usually find?
How does a later layer get its picture?
Who decides the filters in a real CNN?