#channel-mixing

Question 1: Single choice

mediumAI-checkedConvolutional Neural Networks

A grouped convolution with $g$ groups is $g$ times cheaper than a dense one, because its channel-mixing matrix is block-diagonal. What is the main downside, and how does the ResNeXt block reduce it?

The output has only $1/g$ of the channels; concatenating $g$ copies of it restores the full width1
The receptive field shrinks by a factor of $g$ ; larger kernels in later layers make up for it2
The layer needs $g$ times more parameters; sharing weights between the groups removes the excess3
Groups never exchange information; $1 \times 1$ convolutions around it mix the channels again4

Timed test