DenseNet

DenseNet connects every layer to every subsequent layer to maximize feature reuse and eliminate vanishing gradients.

Developed by Gao Huang and his team at Cornell, DenseNet earned the CVPR 2017 Best Paper Award for its radical approach to information flow. By linking each layer to every other layer within a dense block, the architecture ensures that every layer receives collective knowledge from all preceding layers. This design significantly reduces the parameter count compared to ResNet (often by half) while maintaining superior performance on benchmarks like ImageNet and CIFAR-100. It utilizes narrow layers (growth rates as low as 12 or 32) to maintain efficiency, making it a powerhouse for high-accuracy computer vision tasks with limited memory budgets.

https://arxiv.org/abs/1608.06993