Stochastic depth skips some random layers by creating a direct connection. Going one step further, instead of removing some random layers, another way to do the same thing is to add an identity connection with previous layers:
A dense block (densely connected convolutional networks)
As for residual blocks, a densely connected convolutional network consists of repeating dense blocks to create a stack of layer blocks:
A network with dense blocks (densely connected convolutional networks)
Such an architecture choice follows the same principles as those seen in Chapter 10, Predicting Times Sequence with Advanced RNN, with highway networks: the identity connection helps the information to be correctly propagated and back-propagated through the network, reducing the effect of exploding/vanishing gradients when the number of layers is high.
In Python, we replace our residual block with a densely connected block:
def dense_block(network, transition=False, first=False, filters=16): ...
