Visual Computing for Industry, Biomedicine, and Art 2024

Flipover

Flipover Outperforms Dropout in Deep Learning

Yuxuan Liang, Chuang Niu, Pingkun Yan, Ge Wang

Current or former lab member. Select a name to see all of their projects.

Method overview of Flipover
Flipover operations. Left: a standard feedforward network. Middle: the network with dropout. Right: the network with flipover (Fig. 1 of the paper).

Abstract

Dropout regularizes a network by randomly removing neurons and their connections during training. Flipover selects neurons in the same way but, instead of removing them, reverses their outputs with a negative multiplier.

This gives stronger regularization. In experiments across several network types, Flipover matches or exceeds dropout in reducing overfitting, and improves robustness to noise and to adversarial attacks.

What’s in the repository

  • Notebooks for the MNIST and CIFAR-10 experiments, including FGSM attacks and Poisson noise
  • Model definitions

Citation

@article{liang2024flipover,
  title   = {Flipover Outperforms Dropout in Deep Learning},
  author  = {Liang, Yuxuan and Niu, Chuang and Yan, Pingkun and Wang, Ge},
  journal = {Visual Computing for Industry, Biomedicine, and Art},
  volume  = {7},
  pages   = {4},
  year    = {2024},
  doi     = {10.1186/s42492-024-00153-y}
}
Back to projects