“关注人工智能的未来,融合设计与科技”
联系我们: yanyansun@mail.tsinghua.edu.cn

AI为何犯错:解读“机器学习误分类问题“的可视化交互装置设计研究

作品旨在通过直观、有趣的方式,帮助大众理解AI机器学习误分类的原因,达到消解技术鸿沟,更深入地理解人机智能协同的目的。

The work enables the public to intuitively comprehend the reasons behind AI model misclassifications regardless of their level of AI knowledge. It helps bridge the technological gap and improves the collaborative user experience between humans and AI.

Integrating with various AI models, it assists artists and designers in gaining a deeper understanding of how AI makes decisions related to artistic design styles, features, imagery, materials, music rhythm, melody, and chords.

Why AI Errs: Parallax

Parallax aims to visualize the process of AI model transformation from an unexplainable “black box” to a transparent “white box” through a sliding screen. It enables the public to intuitively comprehend the reasons behind AI misclassifications regardless of their level of AI knowledge. It helps bridge the technological gap and improves the collaborative user experience between humans and AI. Users can click on the misclassified images they are interested in on the slide screen to see the primary basis for the classification and compare the differences in data and models between biased and normal AI model. The user experiment proves that the design can make the mass users understand the reasons for AI mistakes.

At Tsinghua University Art Museum, we have carefully designed a structure combining a large 60-inch interactive display with an accompanying 28-inch interactive sliding screen.  The 60-inch screen presents the mysterious AI black boxes, in front of which the 28-inch sliding screen provides an immersive experience showcasing the remarkable visual journey of artificial intelligence, capturing its transformation from black boxes to elucidating white boxes.

Read the paper here:

[X10] Lanxi Xiao, Weikai Yang, Haoze Wang, Shixia Liu and Qiong Wu Why AI Fails: Parallax

Why AI Errs: Shortcut

Shortcut aims to materialize AI data and models by connecting multiple screens with fiber optics and visualize AI models’ shortcuts when making misclassifications. Users can place cards featuring their favorite cats and dogs on the interactive platform and explore why artificial intelligence model may misclassify them, identifying cats as dogs or vice versa.

Watch the video here:

Time
2022-2023
Type
Author
肖岚茜,吴琼,刘世霞,杨维铠

Related Works

  • 联系我们

    • yanyansun@mail.tsinghua.edu.cn
      清华大学美术学院B466
      +86 010 62798891
© 2024 数据与智能创新设计研究所