#inception

Question 1: Select all that apply

d2l's versions of AlexNet and VGG-11 have about 47 and 129 million parameters, while its NiN and GoogLeNet have about 2 and 6 million. Which design choices explain the gap? Select all that apply.

They run on smaller images, and convolution kernels need fewer weights for smaller inputs1
NiN replaces AlexNet's $11 \times 11$ , $5 \times 5$ and $3 \times 3$ convolutions by $1 \times 1$ convolutions throughout2
Their max-pooling layers have fewer weights than the pooling layers of AlexNet and VGG3
Inception blocks cut the channel count with $1 \times 1$ convolutions before the $3 \times 3$ and $5 \times 5$ ones4
NiN and GoogLeNet end with global average pooling instead of 4096-wide fully connected layers5

Timed test