Densely Connected Convolutional Networks

DenseNet keeps earlier feature maps available by concatenating them and asks each layer to add a small new contribution. Follow channel growth and compressed transitions, then separate the paper’s parameter-efficiency evidence from the activation-memory cost of a naive implementation.

Reading focus: How concatenation preserves earlier maps as separate channels while a growth rate limits each new contribution. Why fixed-resolution dense blocks need channel projections and pooling at their boundaries. What the parameter comparison establishes and why it does not guarantee low peak memory or latency.

CVPR 2017. Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. 11 min read, easy difficulty.