The paper introduces the Point Cloud Network (PCN) architecture, a new implementation of linear layers in deep learning networks. The study provides evidence advocating for PCN’s preference over the Multilayer Perceptron (MLP) in linear layers. Models were trained using both MLP and PCN architectures for a direct comparison of linear layers. The key results are model parameter count and top-1 test accuracy over the CIFAR-10 and CIFAR-100 datasets. The PCN equivalent to AlexNet achieved comparable efficacy to the original architecture with a 99.5% reduction of parameters in its linear layers.
Publication date: 25 Sep 2023
Project Page: https://arxiv.org/abs/2309.12996
Paper: https://arxiv.org/pdf/2309.12996
