Scholarlink

<p>This paper presents a novel approach to posterior inference on simulation input parameters that can effectively address the challenge of multi-modality. As simulations become more complex, accounting for the possibility of a multi-modal posterior distribution is crucial. However, traditional posterior inference methods often fall short; their loss functions are structured such that once the simulation outcome approximates the real-world observation, further exploration ceases, leading to a local optimum and a mode-collapsed posterior distribution. To overcome this challenge, we propose a new&#xa0;approach,&#xa0;namely Sequential Neural Joint Estimation, which&#xa0;integrates two significant research directions to enforce further model exploration by maximizing the newly added mutual information loss. Experimental results demonstrate that the proposed approach performs better in multi-modal posteriors, making it a promising method for handling complex simulations that require accurate posterior inference.</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

Sequential Neural Joint Estimation for likelihood-free inference

  • 期刊论文

  • 发表日期：

    2025年5月9日

查看全文
  • Dongjun Kim,
  • Kyungwoo Song,
  • Seungjae Shin,
  • Wanmo Kang,
  • Il-Chul Moon,
  • Weonyoung Joo

摘要

This paper presents a novel approach to posterior inference on simulation input parameters that can effectively address the challenge of multi-modality. As simulations become more complex, accounting for the possibility of a multi-modal posterior distribution is crucial. However, traditional posterior inference methods often fall short; their loss functions are structured such that once the simulation outcome approximates the real-world observation, further exploration ceases, leading to a local optimum and a mode-collapsed posterior distribution. To overcome this challenge, we propose a new approach, namely Sequential Neural Joint Estimation, which integrates two significant research directions to enforce further model exploration by maximizing the newly added mutual information loss. Experimental results demonstrate that the proposed approach performs better in multi-modal posteriors, making it a promising method for handling complex simulations that require accurate posterior inference.