Scholarlink

<p>Prototype-based models are a good candidate for explainable artificial intelligence, and in these models the prototypes store a primary – rough – information about the classes and the classification problem. If <i>backgrounds</i> are defined as <i>anything but the other classes</i>, then we argue that “prototypizing” this background class is redundant: even if there are examples of background data, due to its nature, a thresholding mechanism with a bias value will suffice. To validate the non-prototype background, we use prototype-based learning on medical image classification. Our model has a good performance, whilst maintaining interpretability and – due to the simple wavelet-based prototypes – the reduced computational complexity. In the conclusions, we argue that, by eliminating the prototypes from the background, we have a simpler and more intuitive model.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

On Background Classes in Prototype-Based Architectures

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月22日

查看全文
  • Hanna-Georgina Lieb,
  • Tamás Kaszta,
  • Lehel Csató

摘要

Prototype-based models are a good candidate for explainable artificial intelligence, and in these models the prototypes store a primary – rough – information about the classes and the classification problem. If backgrounds are defined as anything but the other classes, then we argue that “prototypizing” this background class is redundant: even if there are examples of background data, due to its nature, a thresholding mechanism with a bias value will suffice. To validate the non-prototype background, we use prototype-based learning on medical image classification. Our model has a good performance, whilst maintaining interpretability and – due to the simple wavelet-based prototypes – the reduced computational complexity. In the conclusions, we argue that, by eliminating the prototypes from the background, we have a simpler and more intuitive model.