通过系统性评估更好地了解归因方法的差异
IEEE transactions on pattern analysis and machine intelligence
|January 12, 2024
概括
评估深度神经网络的解释性是一项挑战. 这项研究引入了新的方法 (DiFull,ML-Att,AggAtt) 以更公平,更可靠地评估归因技术,改善模型理解.
科学领域:
- 计算机视觉 计算机视觉
- 人工智能的人工智能
- 机器学习 机器学习
背景情况:
- 深度神经网络在视觉任务中表现出色,但由于其"黑子"性质,缺乏可解释性.
- 后期归因方法旨在确定模型决策的有影响力的图像区域,但如果没有基础真相,它们的评估是困难的.
研究的目的:
- 开发新的评估方案,可靠地测量归因方法的忠实性和公平性.
- 为了实现更系统的视觉检查和不同归因技术的比较.
- 在各种模型中研究广泛使用的归因方法的优缺点.
主要方法:
- 引入了DiFull:一种受控的评估设置,通过操纵输入影响来评估归因忠实性.
- 拟议的ML-Att:在相同的网络层上评估所有方法,以确保公平的比较.
- 开发了AggAtt:一种用于对完整数据集的方法进行系统的定性评估的方案.
主要成果:
- 拟议的方案有助于更可靠,更公平地比较归因方法.
- 分析揭示了几种流行的归因技术的优点和缺点.
- 发现后处理平滑步骤显著提高了某些归因方法的性能.
结论:
- 新的评估方案为评估深度神经网络可解释性方法提供了一个强大的框架.
- 更公平,更系统的评估会导致更好地理解和选择归因技术.
- 提出的方法和后处理步骤有助于推进可解释AI领域.
更多相关视频
08:24The Joint Effect of Social Comparison and Social Distance on Evaluation of Intertemporal Choice Outcomes in Event-related Potential Studies
Published on: August 25, 2023
734
08:27Applying an eMASS Customization Program as a Research Tool to Evaluate Consumer Benefits
Published on: September 27, 2019
6.9K
相关概念视频
Fundamental Attribution Error
12.9K
According to some social psychologists, people tend to overemphasize internal factors as explanations—or attributions—for the behavior of other people. They tend to assume that the behavior of another person is a trait of that person, and to underestimate the power of the situation on the behavior of others. They tend to fail to recognize when the behavior of another is due to situational variables, and thus to the person’s state. This erroneous assumption is...
12.9K
Attribution Theory
13.0K
Behavior is a product of both the situation (e.g., cultural influences, social roles, and the presence of bystanders) and of the person (e.g., personality characteristics). Subfields of psychology tend to focus on one influence or behavior over others. Situationism is the view that our behavior and actions are determined by our immediate environment and surroundings. In contrast, dispositionism holds that our behavior is determined by internal factors (Heider, 1958).
13.0K
Self-Presentation: Self-Monitoring and Self-Handicapping
39.0K
People can go to great lengths to protect their self-image and present themselves in ways that they want others to see them. Sociologist Erving Goffman presented the idea that a person is like an actor on a stage. Calling his theory dramaturgy, Goffman believed that we use “impression management” to present ourselves to others as we hope to be perceived. Each situation is a new scene, and individuals perform different roles depending on who is present (Goffman, 1959). Think about...
39.0K
Group Design
8.9K
The most basic experimental design involves two groups: the experimental group and the control group. The two groups are designed to be the same except for one difference— experimental manipulation. The experimental group gets the experimental manipulation—that is, the treatment or variable being tested—and the control group does not. Since experimental manipulation is the only difference between the experimental and control groups, we can be sure that any differences between...
8.9K
Naturalistic Observations
15.4K
If you want to understand how behavior occurs, one of the best ways to gain information is to simply observe the behavior in its natural context. However, people might change their behavior in unexpected ways if they know they are being observed. How do researchers obtain accurate information when people tend to hide their natural behavior? As an example, imagine that your professor asks everyone in your class to raise their hand if they always wash their hands after using the restroom. Chances...
15.4K
Stereotype Content Model
14.7K
The Stereotype Content Model (SCM) was first proposed by Susan Fiske and her colleagues (Fiske, Cuddy, Glick & Xu, 2002; see also Fiske, 2012 and Fiske, 2017). The SCM specifies that when someone encounters a new group, they will stereotype them based on two metrics: warmth—or that group’s perceived intent, and how likely they are to provide help or inflict harm—and competence—or their ability to carry out that objective. Depending on the warmth-competence...
14.7K
