Scholarlink

<p>Generative artificial intelligence (GenAI) has transformed image synthesis, yet using these models to construct images that accurately represent a human mental image remains a challenge. Current “prompting” methods rely on explicit verbal descriptions, which can fail to capture the complex nuances of a mental representation. <i>BrainPrompting</i> is a neuroadaptive method that utilises implicit brain activity to navigate the latent space of a generative model. Candidate stimuli generated from a generative neural network model are presented to ’query’ users, who ’reply’ via implicit, EEG-based recognition signals. Previous work within this field demonstrated the capability of generating images matching mental categories and visualising individual differences, yet it remains uncertain how well generated images match human representations. To investigate, a “police line-up” experiment was conducted in which target faces were kept in mind while a stream of faces similar or dissimilar to the target were shown. Faces that were reconstructed using BrainPrompting were shown as closely resembling human memory, demonstrating strong alignment between the neural network model and mental representation. By showing that this accuracy increases as a function of aggregating neural signals across multiple users, we present evidence for an above-individual level of accuracy.</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

BrainPrompting: reconstructing facial memory by brain-based generative AI interaction

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月1日

查看全文
  • Aotong Li,
  • Tuukka Ruotsalo,
  • Jian Hwee Ang,
  • Michiel M. Spapé

摘要

Generative artificial intelligence (GenAI) has transformed image synthesis, yet using these models to construct images that accurately represent a human mental image remains a challenge. Current “prompting” methods rely on explicit verbal descriptions, which can fail to capture the complex nuances of a mental representation. BrainPrompting is a neuroadaptive method that utilises implicit brain activity to navigate the latent space of a generative model. Candidate stimuli generated from a generative neural network model are presented to ’query’ users, who ’reply’ via implicit, EEG-based recognition signals. Previous work within this field demonstrated the capability of generating images matching mental categories and visualising individual differences, yet it remains uncertain how well generated images match human representations. To investigate, a “police line-up” experiment was conducted in which target faces were kept in mind while a stream of faces similar or dissimilar to the target were shown. Faces that were reconstructed using BrainPrompting were shown as closely resembling human memory, demonstrating strong alignment between the neural network model and mental representation. By showing that this accuracy increases as a function of aggregating neural signals across multiple users, we present evidence for an above-individual level of accuracy.