Scholarlink

<p>Adaptive neural networks offer a promising approach to reduce the computational cost of neural network inference. These models dynamically allocate computational resources based on each individual input, leading to a significant reduction in average computational cost. In this paper, we show that adaptive computation can have unintended consequences, potentially leading to a significantly worse user experience for certain groups. We train several state-of-the-art adaptive neural networks for the task of age prediction on the FairFace dataset and demonstrate that the average computational cost varies considerably across different user demographics, including age, race and gender. Additionally, we analyze the properties of images that influence the computational cost and show that these factors partially explain the observed demographic differences in performance.</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

Computational fairness in adaptive neural networks

  • 期刊论文

  • 发表日期：

    2025年6月30日

查看全文
  • Sam Leroux,
  • Ciem Cornelissen,
  • Vishisht Sharma,
  • Pieter Simoens

摘要

Adaptive neural networks offer a promising approach to reduce the computational cost of neural network inference. These models dynamically allocate computational resources based on each individual input, leading to a significant reduction in average computational cost. In this paper, we show that adaptive computation can have unintended consequences, potentially leading to a significantly worse user experience for certain groups. We train several state-of-the-art adaptive neural networks for the task of age prediction on the FairFace dataset and demonstrate that the average computational cost varies considerably across different user demographics, including age, race and gender. Additionally, we analyze the properties of images that influence the computational cost and show that these factors partially explain the observed demographic differences in performance.