Visual Computing for Industry, Biomedicine, and Art 2022

QuadraticNeurons

Superiority of Quadratic over Conventional Neural Networks for Classification of Gaussian Mixture Data

Tianrui Qi, Ge Wang

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

Method overview of QuadraticNeurons
The affine layer of a conventional neuron (left) and a quadratic neuron (right) (Fig. 3 of the paper).

Abstract

A quadratic neuron replaces the inner product between inputs and weights with a quadratic operation, which adds diversity to the building blocks of a network. This study trains quadratic and conventional networks with an adapted backpropagation algorithm and compares them on the classification of Gaussian mixture data.

In this setting the quadratic networks perform remarkably better than conventional networks in both efficacy and efficiency.

What’s in the repository

  • Implementations of the neural networks, the Gaussian mixture data generator and an expectation-maximization baseline
  • Notebooks for the illustrative and systematic experiments and for three real datasets (ISOLET, Pendigits, Yeast)

Citation

@article{qi2022superiority,
  title   = {Superiority of Quadratic over Conventional Neural Networks for Classification of Gaussian Mixture Data},
  author  = {Qi, Tianrui and Wang, Ge},
  journal = {Visual Computing for Industry, Biomedicine, and Art},
  volume  = {5},
  pages   = {23},
  year    = {2022},
  doi     = {10.1186/s42492-022-00118-z}
}
Back to projects