FluxNet is a novel neural architecture for computer vision that incorporates stateful processing, adaptive gateway routing, and iterative refinement mechanisms. It introduces stateful feature processing, adaptive channel gateway for dynamic feature routing, and iterative refinement blocks for multi-step feature refinement with state feedback. FluxNet achieves superior accuracy on CIFAR-10 and CIFAR-100 benchmarks compared to ResNet-50 and Vision Transformer baselines, while maintaining comparable inference efficiency and exhibiting robustness to common corruptions.
Key findings
FluxNet achieves 96.38% accuracy on CIFAR-10 and 81.15% on CIFAR-100, outperforming ResNet-50 and ViT-B/16.
Stateful processing provides the largest individual gain of +0.62% on CIFAR-10.
FluxNet exhibits superior robustness to common corruptions, with a mean accuracy drop of only 1.52% compared to 1.92% for ResNet-50 and 2.45% for ViT-B/16.
Limitations & open questions
FluxNet's performance on larger and more complex datasets remains to be evaluated.