[2021] Generative Classifiers as a Basis for Trustworthy Image Classification

[gatheluck]

論文リンク

• CVF

公開日（yyyy/mm/dd）

CVPR 2021

概要

TeX

% yyyy/mm/dd
@inproceedings{
    mackowiak2021generative,
    title={Generative Classifiers as a Basis for Trustworthy Image Classification},
    author={Radek Mackowiak and Lynton Ardizzone and Ullrich Kothe and Carsten Rother},
    booktitle=CVPR,
    year={2021}
}

[gatheluck]

• they define " trustworthiness" ~ "the combination of explainability and robustness"
• Generative classifiers (GCs) are a promising class of models that naturally accomplish these.
• contribution
  • test GCs to more complex task like ImageNet
  • show trustworthiness of GCs
• claim: GCs are a highly promising basis for further algorithms and modifications.

[gatheluck]

Invertible neural networks (INNs)

• the network represents a diffeomorphism by construction (essentially, a smooth and invertible function),
• the inversion can be computed efficiently
• the network has a tractable Jacobian determinant.

[gatheluck]

problem

• X: training images
• p(X): unknown image distribution
• q_{theta}(X): distribution given by the network
• goal: approximate p(X) as closely as possible with q_{theta}(X)