Academia.eduAcademia.edu

Figure 2 – uploaded by Sebastian Palacio

See full PDFdownloadDownload figure
Table 2. Definitions of “explanation” and “interpretation” according to various dictionaries. Explanation is referred to as an object while interpretation is more commonly associated to an action. Accessed on 05.01.2021. requirements. Asking why something happens, inescapably relates back to causal effects. While these kind of relationships are among the most useful to discover (as it allows for more control over the effect by adjusting the cause), other non-causal relationships remain valuable in the toolset of explainability. Finding out that a model is unfair, without knowing what the cause of it is, can already be helpful in high-stake sce- narios (e.g., by preventing its use). We say that the why relates to the nature of an interpretation. In short, if the interpretation bares a causal meaning, then the why is being defined by the causal link. If the meaning is limited to a correlation, the why is left out of the scope of that particular interpretation. Note that, if the explanation method is al- ready based on causal theory (Lopez-Paz et al., 2017; Chang et al., 2019), the assigned meaning (i.e., the link between the explanans and the high-level (non-)functional requirement) will be more direct and therefore, more likely to withstand scientific scrutiny. The why is therefore not mandatory in explanations generated by XAI methods. In any case, the explanation’s context can and should be defined, be it causal or based solely on correlations. Proponents of XAI methods are responsible for clearly stating the context in which their explanations can be interpreted.

Table 2 Definitions of “explanation” and “interpretation” according to various dictionaries. Explanation is referred to as an object while interpretation is more commonly associated to an action. Accessed on 05.01.2021. requirements. Asking why something happens, inescapably relates back to causal effects. While these kind of relationships are among the most useful to discover (as it allows for more control over the effect by adjusting the cause), other non-causal relationships remain valuable in the toolset of explainability. Finding out that a model is unfair, without knowing what the cause of it is, can already be helpful in high-stake sce- narios (e.g., by preventing its use). We say that the why relates to the nature of an interpretation. In short, if the interpretation bares a causal meaning, then the why is being defined by the causal link. If the meaning is limited to a correlation, the why is left out of the scope of that particular interpretation. Note that, if the explanation method is al- ready based on causal theory (Lopez-Paz et al., 2017; Chang et al., 2019), the assigned meaning (i.e., the link between the explanans and the high-level (non-)functional requirement) will be more direct and therefore, more likely to withstand scientific scrutiny. The why is therefore not mandatory in explanations generated by XAI methods. In any case, the explanation’s context can and should be defined, be it causal or based solely on correlations. Proponents of XAI methods are responsible for clearly stating the context in which their explanations can be interpreted.

Related Figures (3)

XAI Handbook: Towards a Unified Framework for Explainable AI Figure 1. Overview of our proposed framework. A task defined in the high-level domain gets an under-specified characterization, leaving out non-functional requirements. “Explanations” are meth- ods that probe for said requirements. Interpretations are mappings from low- to the high-level domain.
XAI Handbook: Towards a Unified Framework for Explainable AI Figure 1. Overview of our proposed framework. A task defined in the high-level domain gets an under-specified characterization, leaving out non-functional requirements. “Explanations” are meth- ods that probe for said requirements. Interpretations are mappings from low- to the high-level domain.
to ask what kind of semantic entity an explanation is: is it an action, an outcome, a process or an object? For most textbook definitions (see Table 2), explanations are seen as statements. In turn, such statements are nothing but descrip- tions about an already existing entity (the explanandum or the one which is subject to description). From a functional perspective, an explanation is therefore the process by which an explanandum is described. Finally, to avoid confusions between the process of explaining and its output, we refer to the former as explanation and the latter as the explanans. As we seek explanations for AI models, we find such suit- able facts in the form of low-level mathematical primitives used to build the models themselves: support vector ma- chines have a decision boundary equation, coordinates of the support vectors, the tolerance threshold, etc. A Neural Network has values for each individual parameter, the equa- tions governing how they connect with each other, the value of the cost function when computed on a particular input, the gradient that can be computed on that loss, etc. All of these are concrete, undisputed facts (assuming there are no bugs) that are suitable explanandums.
to ask what kind of semantic entity an explanation is: is it an action, an outcome, a process or an object? For most textbook definitions (see Table 2), explanations are seen as statements. In turn, such statements are nothing but descrip- tions about an already existing entity (the explanandum or the one which is subject to description). From a functional perspective, an explanation is therefore the process by which an explanandum is described. Finally, to avoid confusions between the process of explaining and its output, we refer to the former as explanation and the latter as the explanans. As we seek explanations for AI models, we find such suit- able facts in the form of low-level mathematical primitives used to build the models themselves: support vector ma- chines have a decision boundary equation, coordinates of the support vectors, the tolerance threshold, etc. A Neural Network has values for each individual parameter, the equa- tions governing how they connect with each other, the value of the cost function when computed on a particular input, the gradient that can be computed on that loss, etc. All of these are concrete, undisputed facts (assuming there are no bugs) that are suitable explanandums.
Table 3. Recontextualization of three popular XAI methods with the help of our proposed framework.
Table 3. Recontextualization of three popular XAI methods with the help of our proposed framework.
Related topics:
Computer Science
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved