1. Though (a) is the idea of ResNeXt to achieve higher cardinality, (b) from inception-ResNet and (c) from AlexNet can also achieve the high cardinality but with a bit different structure. Finally, authors uses AlexNet implementation for the grouped convolution. (In AlexNet, only C=2 is used, and the purpose is to distribute the number of operations in two GPUs. In ResNeXt, C can be larger than 2, and is to boost the accuracy.)

2. should be 64 conv paths <<< Thanks for pointing out. Corrected.