XnODR and XnIDR: Two Accurate and Fast Fully Connected Layers for Convolutional Neural Networks

Keywords:

CapsNet, XNOR-Net, Dynamic routing, Binarization, Xnorization, Machine learning, Neural network.

Contributions:

• We propose a new Fully Connected Layer called XnODR by deploying Xnorization on the CapsFC layer. To be specific, we xnorize the linear projection outside the dynamic routing.
• We propose a new Fully Connected Layer called XnIDR by xnorizing another linear projection inside the dynamic routing.
• XnODR and XnIDR improve the performance (i.e., better accuracy, less FLOPS, and parameters) of both lightweight (MobileNetV2) and heavyweight (ResNet50) models.

Production: XnODR and XnIDR

Used Framework:

BibTex:

@article{sun2023xnodr,
  title={Xnodr and xnidr: Two accurate and fast fully connected layers for convolutional neural networks},
  author={Sun, Jian and Fard, Ali Pourramezan and Mahoor, Mohammad H},
  journal={Journal of Intelligent \& Robotic Systems},
  volume={109},
  number={1},
  pages={17},
  year={2023},
  publisher={Springer}
}

Recommended citation: Sun, J., Fard, A.P. & Mahoor, M.H. XnODR and XnIDR: Two Accurate and Fast Fully Connected Layers for Convolutional Neural Networks. J Intell Robot Syst 109, 17 (2023).
Download Paper | Download Slides