11-19 更新 实时
首页/ 算力/ 正文

新 AI 方法揭示视觉模型识别哪些概念及位置

算力 来源:TechXplore 2026-11-18 23:40 0 阅读 机器翻译
新 AI 方法揭示视觉模型识别哪些概念及位置

继续使用 Google、Facebook 或 Apple,即表示你同意使用条款。

新 AI 方法揭示视觉模型识别出哪些概念,以及这些概念出现在何处

作者:Philipp Zapf-Schramm,马克斯·普朗克学会

编辑:Lisa Lock,审阅:Andrew Zinin

本文已按照 Science X 的编辑流程与政策进行审核。编辑在确保内容可信度的同时,重点标注了以下属性:

添加到首选来源

基础模型是在海量数据上训练的 AI 模型,随后可被适配以执行各种各样的任务——就像一套支持多种程序的操作系统。其中一类是语言对齐的视觉基础模型(VLM),它在数百万对图像-文本上训练,因而能够把图像与语言关联起来。如今,它们的应用范围十分广泛,从图像识别、图像搜索到图像生成,还会与大型语言模型结合使用。

然而,这些模型有一个关键弱点:人类不易理解它们是如何得出结果的。例如,模型可能在一张医学图像中正确识别出恶性痣——但直到现在,仍很难确定模型为何识别出它,以及它在图像中实际“看到”了什么。这种透明度的缺失,被认为是 AI 系统大规模推广的主要障碍,尤其是在医学等安全关键或敏感领域。

“只有当我们理解一个模型的决策基于什么,才能检验它是否正确是出于正确的原因,或者在它犯错时找出它为何失败,”位于萨尔布吕肯的马克斯·普朗克信息学研究所可解释机器学习研究组负责人 Jonas Fischer 博士说。

这正是由 Kai Wittenmayer、Sukrut Rao、Amin Parchami-Araghi、Bernt Schiele 和 Fischer 组成的研究团队所切入的方向——他们提出了一种新的建模方法:面向视觉的语言对齐概念基础模型(Language-aligned Concept Foundation Model for Vision,CFM)。该团队在瑞典马尔默举行的欧洲计算机视觉会议(ECCV 2026)上展示了他们的工作。

首先,一张图像被划分为称为图块(patch)的小区域。一个小型网络经过专门训练,把模型中表示的数值(向量数据)转换为明确定义的概念,并为它们赋予简洁、人类可读的标签。

由于每个图块都被单独分析,系统能够确定每个概念出现在图像中的什么位置。通过识别在同一区域共同出现的概念,该方法还能建立更宽泛和更细分的类别——例如,“帆船”是“船/舟”的一种具体类型。

与此前只是列出模型似乎在一张图像中识别出的概念的方法不同,CFM 通过精确定位各个概念在图像中出现的位置来解释其决策。

“CFM 不仅能揭示模型在一张图像中识别出了哪些概念,还能有选择地对它们施加影响。例如在图像描述任务中,可以抑制个别概念,使某些图像内容不再出现在图像描述中,而其他内容则继续被描述,”论文第一作者、博士生 Wittenmayer 解释道。

研究人员的这项概念验证研究开辟了广泛的应用可能:语言对齐的视觉基础模型往往为大型语言模型(LLMs)同时处理图像与文本提供了基础。新的概念基础模型(Concept Foundation Model)能够以人类可理解的方式拆解并解释图像的各个组成部分。这样一来,人们可以在早期就识别并纠正模型不可靠或错误的行为,而无需从头重新训练模型。这有望省去耗时且资源密集的训练,从而提升效率。

CFM 是首个基于可理解的概念、为多种下游任务提供解释的基础模型,这些任务包括图像分类、图像描述和图像分割。此前的可解释性方法大多局限于图像分类。额外的解释层对性能只有极小影响。在测试中,CFM 在上述各项任务上的性能水平与作为其基础的不透明模型几乎一致。

“凡是需要 AI 决策可解释的场景,CFM 都可能派上用场。例如,它最终可能被用于医学领域,以确保黑色素瘤是依据诊断特征被识别出来的,而不是因为图像中出现了医院手套,”马克斯·普朗克信息学研究所(Max Planck Institute for Informatics)计算机视觉与机器学习部门主任 Schiele 解释道。接下来,研究人员计划将他们的概念表示与大型语言模型和图像模型相结合,例如用于视觉问答系统。

论文:Kai Wittenmayer 等,《CFM: Language-aligned Concept Foundation Model for Vision》,在瑞典马尔默举行的欧洲计算机视觉会议(ECCV 2026)上发表

这篇报道的背后是谁?

艺术史学士,物质文化硕士。曾任博物馆编辑、急救员和移植协调员。自 2021 年起在 Science X 担任编辑。完整简介 →

物理学硕士,有科研经历。长期的科学新闻爱好者。在 Science X 的编辑工作中发挥关键作用。完整简介 →

引用:新 AI 方法揭示视觉模型识别哪些概念以及这些概念出现在何处(2026年11月18日),2026年11月18日检索自 https://techxplore.com/news/2026-11-ai-method-reveals-concepts-vision.html

本文件受版权保护。除为私人学习或研究目的进行的合理使用外,未经书面许可不得复制任何部分。内容仅供参考。

可解释 AI:新框架提升决策系统的透明度

开源框架超越语言,增强多模态 AI 训练能力

用于生物标志物检测的数据高效基础模型

研究提出可解释 AI 的新方法

用插画训练无图像计算机视觉系统识别真实照片

帮助 AI 文本检测器跟上不断变化的环境

让氢火焰更易被看见,或有助于推动更清洁的能源未来

电子指纹帮助 AI 在极少训练数据下预测复杂合金特性

机器人学习人类运动模式,在密集人群中找到通行路径

这种微型磁体可将普通晶体管电路变为人工神经元

单螺旋桨无人机利用自身旋转来全方位扫描周围环境

获取即时摘要文本(Gist)

Concept Foundation Model 将语言对齐的视觉模型表征映射为人类可读的概念,并在图像中定位这些概念。它能以极小的性能损失解释分类、图像描述和分割。选择性概念抑制会改变图像描述,同时保留对其他内容的提及。

本摘要由 LLM 自动生成。完整免责声明

译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(New AI method reveals which concepts vision models recognize and where they appear)

By continuing with Google, Facebook or Apple, you agree to the Terms of Use.

New AI method reveals which concepts vision models recognize and where they appear

by Philipp Zapf-Schramm, Max Planck Society

edited by Lisa Lock, reviewed by Andrew Zinin

This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility:

Add to Preferred Sources

New AI method reveals which concepts vision models recognize and where they appear Credit: Max Planck Society

Foundation models are AI models trained on enormous amounts of data that can subsequently be adapted to perform a wide range of tasks—much like an operating system that supports a variety of programs. One type is the language-aligned vision foundation model (VLM), which is trained on millions of image-text pairs, enabling it to connect images with language. Today, they are used in a wide range of applications, from image recognition and image search to image generation, and in combination with large language models.

However, these models have a critical weakness: Humans cannot easily understand how they arrive at their results. For example, a model might correctly identify a malignant mole in a medical image—but until now, it has been difficult to determine why the model identified it and what it actually "saw" in the image. This lack of transparency is considered a major obstacle to the widespread adoption of AI-based systems, especially in safety-critical or sensitive areas such as medicine.

"Only if we understand what a model's decision is based on can we check whether it is correct for the right reasons, or find out why it failed when it made a mistake," says Dr. Jonas Fischer, head of the research group Explainable Machine Learning at the Max Planck Institute for Informatics in Saarbrücken.

This is where the research team consisting of Kai Wittenmayer, Sukrut Rao, Amin Parchami-Araghi, Bernt Schiele and Fischer comes in—with a new modeling approach, the Language-aligned Concept Foundation Model for Vision (CFM). The team presented their work at the European Conference on Computer Vision (ECCV 2026), in Malmö, Sweden.

First, an image is divided into small sections called patches. A small network is specifically trained to translate the numerical values (vector data) represented in the model into clearly defined concepts and assign them concise, human-readable labels.

Because each patch is analyzed individually, the system can determine where each concept appears in the image. By identifying concepts that occur together in the same regions, the method can also establish broader and narrower categories—for example, that a "sailboat" is a specific type of "ship/boat."

Unlike previous approaches that simply list the concepts a model appears to recognize in an image, CFM explains its decisions by pinpointing where the individual concepts appear in the image.

"CFM can not only reveal which concepts a model recognizes in an image, but also selectively influence them. In image description tasks, for example, individual concepts can be suppressed so that certain image contents are no longer mentioned in the image description, while others continue to be described," explains doctoral student Wittenmayer, the paper's first author.

The researchers' proof-of-concept study opens up far-reaching application possibilities: Language-aligned vision foundation models often provide the foundation that enables large language models (LLMs) to process images and text simultaneously. The new Concept Foundation Model makes it possible to break down and explain image components in a way that humans can understand. This makes it possible to identify and correct unreliable or erroneous model behavior early on, without having to retrain a model from scratch. This could improve efficiency by eliminating time-consuming and resource-intensive training.

CFM is the first foundation model that provides explanations for a variety of downstream tasks based on understandable concepts, including image classification, image description and image segmentation. Previous approaches to explainability have mostly been limited to image classification. The additional explanatory layer has only a minimal impact on performance. In tests, CFM achieved performance levels nearly identical to those of the underlying, opaque models across the tasks mentioned above.

"CFM could be useful wherever AI decisions need to be explainable. For example, it could eventually be used in medicine, to ensure that a melanoma was identified based on diagnostic features, rather than the presence of hospital gloves in the image," explains Schiele, director of the Computer Vision and Machine Learning department at the Max Planck Institute for Informatics. Next, the researchers plan to combine their concept representation with large language and image models, for example for use in visual question-answering systems.

Paper: Kai Wittenmayer et al., CFM: Language-aligned Concept Foundation Model for Vision, presented at the European Conference on Computer Vision (ECCV 2026), Malmö, Sweden

Who's behind this story?

Lisa Lock

BA art history, MA material culture. Former museum editor, paramedic, and transplant coordinator. Editing for Science X since 2021. Full profile →

Master's in physics with research experience. Long-time science news enthusiast. Plays key role in Science X's editorial success. Full profile →

Citation: New AI method reveals which concepts vision models recognize and where they appear (2026, November 18) retrieved 18 November 2026 from https://techxplore.com/news/2026-11-ai-method-reveals-concepts-vision.html

This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no part may be reproduced without the written permission. The content is provided for information purposes only.

Explainable AI: New framework increases transparency in decision-making systems

Open-source framework goes beyond language to enhance multimodal AI training capabilities

A data-efficient foundation model for biomarker detection

Study presents new method for explainable AI

Using illustrations to train an image-free computer vision system to recognize real photos

Helping AI text detectors keep up with changing conditions

Making hydrogen flames easier to see could help power a cleaner energy future

Electron fingerprints help AI predict complex alloy properties with little training data

Robots learn human movement patterns to find paths through dense crowds

This tiny magnet could turn ordinary transistor circuits into artificial neurons

Single-propeller drone uses its own spin to scan surroundings in all directions

Get Instant Summarized Text (Gist)

The Concept Foundation Model maps language-aligned vision model representations to human-readable concepts and localizes them within images. It explains classification, captioning and segmentation with minimal performance loss. Selective concept suppression alters image descriptions while preserving mentions of other content.

This summary was automatically generated using LLM. Full disclaimer

This site uses cookies to assist with navigation, analyse your use of our services, collect data for ads personalisation and provide content from third parties. By using our site, you acknowledge that you have read and understand our Privacy Policy and Terms of Use.

原文为英文,正文由自动抓取并保留原文。

本文由采集器自动同步

原文发布于 TechXplore(New AI method reveals which concepts vision models recognize and where they appear),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗