All convolutions in the dense block are ReLU-activated and use batch normalization. Channel-sensible concatenation is barely possible if the peak and width dimensions of the data keep on being unchanged, so convolutions in a very dense block are all of stride one. Pooling levels are inserted between dense blocks https://financefeeds.com/zoth-launches-first-ever-rwa-restaking-layer-with-zeusd-announces-exclusive-pre-deposit-campaign/