#densenet

Question 1: Single choice

d2l builds a DenseNet with four dense blocks of four convolution blocks each (growth rate 32), to match the four stages of its ResNet-18. The DenseNet has about 0.76 million parameters and the ResNet-18 about 11.2 million. What is the main reason?

DenseNet replaces batch normalization, which holds many parameters, with average pooling in every block1
DenseNet uses $1 \times 1$ kernels in its dense blocks, while ResNet-18 uses $3 \times 3$ kernels throughout2
Concatenation shares weights between layers, so each kernel has to be stored only once for the whole network3
Each layer adds only 32 channels and gets earlier features by concatenation, so its convolutions stay narrow4

Timed test