# DenseNet Building on [[ResNet]], DenseNets make the following improvements: - Add skip connections to multiple forward layers, (not just consequent layers) $ y=h\left(x_{l}, x_{l-1}, \ldots, x_{l-n}\right) $ - Assume layer 1 captures edges, while layer 5 captures faces (and other stuff) - Why not have a layer that combines both faces and edges (e.g. to model a scarred face) - Standard ConvNets do not allow for this,exL Layer 6 combines only layer 5 patterns, not lower. ![[densenet.png]] --- ## References 1. Lecture 5.4, UvA DL course 2020