Part of International Conference on Representation Learning 2025 (ICLR 2025) Conference
Ashish Kumar, Jaesik Park
In the era of vision Transformers, the recent success of VanillaNet shows the hugepotential of simple and concise convolutional neural networks (ConvNets). Wheresuch models mainly focus on runtime, it is also crucial to simultaneously focuson other aspects, e.g., FLOPs, parameters, etc, to strengthen their utility further.To this end, we introduce a refreshing ConvNet macro design called ColumnarStage Network (CoSNet). CoSNet has a systematically developed simple andconcise structure, smaller depth, low parameter count, low FLOPs, and attention-less operations, well suited for resource-constrained deployment. The key noveltyof CoSNet is deploying parallel convolutions with fewer kernels fed by inputreplication, using columnar stacking of these convolutions, and minimizing the useof 1×1 convolution layers. Our comprehensive evaluations show that CoSNet rivalsmany renowned ConvNets and Transformer designs under resource-constrainedscenarios. Pretrained models shall be open-sourced.