architectureIntroduced by ResNeXt · 2016
Cardinality (grouped convolutions)
Split a block into many parallel same-shaped paths; the number of paths is a new scaling axis.
Drafted by AI · not yet reviewed
How this idea evolved
Suggested from citations. No curator has recorded what this idea builds on yet. These ideas from the same theme come from papers that ResNeXt cites, directly or one step removed. Citation is a fact; the connection between the ideas is not verified.
Add a layer's input to its output so very deep networks stay trainable.
Parallel convolutions of several sizes in one block, to go deeper and wider at fixed cost.
Stack many small 3×3 convolutions; depth alone drives accuracy.