DPNs

DPNs, or Dual Path Networks, are a family of convolutional neural networks designed to combine the strengths of residual networks (ResNet) and dense networks (DenseNet) for image recognition. The core idea is to enable both feature reuse and new feature exploration within each block by maintaining two parallel pathways: a residual path that uses additive fusion and a dense path that uses concatenation to produce new features. Within a block, the input feature maps are processed to yield two sets of features which are then merged, resulting in an output whose channel dimension grows with depth.

The dual-path structure aims to preserve gradient flow while encouraging dense connectivity. This combination seeks to deliver improved learning efficiency and accuracy without a prohibitive increase in parameter count. Variants of DPNs, such as DPN-68 and DPN-92, differ in depth and width to suit different computational budgets. DPNs have been evaluated on standard benchmarks such as ImageNet and CIFAR, where they demonstrate competitive accuracy with relatively efficient parameter usage compared with purely residual or dense architectures.

Limitations include higher memory usage due to concatenated feature maps and added architectural complexity that can complicate deployment. DPNs have influenced subsequent work on hybrid connectivity in neural networks and are cited as an example of integrating residual and densely connected design principles to balance feature reuse with exploration of new representations.