The Pragmatic Turn in Explainable Artificial Intelligence (XAI)

Abstract

In this paper I argue that the search for explainable models and interpretable decisions in AI must be reformulated in terms of the broader project of offering a pragmatic and naturalistic account of understanding in AI. Intuitively, the purpose of providing an explanation of a model or a decision is to make it understandable to its stakeholders. But without a previous grasp of what it means to say that an agent understands a model or a decision, the explanatory strategies will lack a well-defined goal. Aside from providing a clearer objective for XAI, focusing on understanding also allows us to relax the factivity condition on explanation, which is impossible to fulfill in many machine learning models, and to focus instead on the pragmatic conditions that determine the best fit between a model and the methods and devices deployed to understand it. After an examination of the different types of understanding discussed in the philosophical and psychological literature, I conclude that interpretative or approximation models not only provide the best way to achieve the objectual understanding of a machine learning model, but are also a necessary condition to achieve post-hoc interpretability. This conclusion is partly based on the shortcomings of the purely functionalist approach to post-hoc interpretability that seems to be predominant in most recent literature.

What is Interpretability?

“…Since we argue in Section 3 that ANNs are explainable using the traditional models described above, it may be argued that our account of interpretability could inherit some of these issues. We agree that the causal-interventionist account may be useful for solving some questions about explainability, understandability, and interpretability in AI (Páez 2019). Indeed, we describe interpretability processes in Section 4.3 some of which result in understandable explanations which include counterfactual information, however, the causal-interventionist account includes pragmatic elements that we maintain should not count against a phenomenon's explainability, namely that an explanation should be "deeper or more satisfying" (Woodward 2003: 190) than those provided by DN, IS, or CM explanations.…”

“…(Lipton 2018) We contend that much confusion in the debate about, and push for, XAI can be attributed to a conflation of explainability and interpretability. Interpretation has been devised and variously defined within the sciences themselves (Ribeiro et al 2016;Mittelstadt et al 2019;Lipton 2018;Krishnan 2019;Páez 2019). We cannot fault philosophers or scientists for misunderstanding the scientific notion of interpretation since there is no single such notion to rely on.…”

“…And in that somewhat trivializing case, an explanation of the classification of the input can be reduced to a selection of features of the input that most affect the output; such local interpretations provide counterfactual information about which features of the input, if intervened upon, would change the output. For example, local interpretation methods in MAIS often identify the particular pixels, present in an input image, that most affect the classification output (Páez 2019;Zednik 2019; see also Ribeiro et al 2016). In practice, methods for interpretation of predictions of machine learning algorithms often approximate the explanation, in addition to localizing it to a given datum, as examined in the following section.…”

The explanation game: a formal framework for interpretable machine learning

“…Pragmatic goals require pragmatic strategies. Because iML is fundamentally about getting humans to understand the behaviour of machines, there is a growing call for personalised solutions (Páez 2019). We take this pragmatic turn seriously and propose formal methods to implement it.…”

The State Space of Artificial Intelligence

“…The notion of self-x-capacity is related to the program of organic computing, where the notion of self-x-property is standardly used (self-repairing 1 The recent developments in AI only start to get the full attention from philosophy of science and philosophy of mind and cognition that they deserve: cf. Buckner (2018Buckner ( , 2019, López-Rubio (2018), Páez (2019), Schubbach and Arno (2019), and Zednik (2019). 2 I chose the term "AI state space", or sometimes just "AI space" for short.…”