Aleks Knoks

Deep Neural Networks and other AI systems engineered using advanced machine learning techniques can tackle a wide range of tasks with proficiency that seems to match and even surpass human ability. Yet they are also notoriously opaque, and the worries surrounding their opacity have given rise to the burgeoning field of explainable artificial intelligence, or XAI, with its large variety of explainability methods. This includes post-hoc explainability methods which purport to explain opaque AI systems on the basis of their input-output behaviour alone, without looking into their inner workings. But many authors have expressed deep scepticism about post-hoc explainability. In this paper, we explore the two main objections that the sceptics have raised and argue that we shouldn’t be swayed by them. Using ideas from recent literature on scientific explanation, we draw a parallel between post-hoc explainability methods and idealized models in science and use the parallel in our reply to the sceptics.

The informal literature on normative reasons standardly accepts two claims: that the way reasons compete when determining oughts can be made sense of with the help of the metaphor of balancing or weighing reasons on a scale; and that, in addition to reasons for, there are also reasons against. However, Justin Snedegar (Phil Studies, 2018) has recently argued that these claims are in tension, and that this tension suggests that the balancing metaphor has to be given up. This paper approaches reasons against and the tension, drawing on a defeasible logic-based model of the way reasons compete when determining oughts. The formal exploration points to a natural way to resolve the tension that works not only in the context of the formal model, but also the standard informal view

Intuitively, one can do more than morality requires. Thus, you seem to do more than morality requires---in jargon, you supererogate---when you knowingly let grave harm come to you to save a stranger's life. But while supererogation is part and parcel of our commonsense morality, it is also a puzzling phenomenon. Among other things, it gives rise to three puzzles. This paper develops a formal system, or logic, that resolves these three puzzles. The system emerges once we combine a well-known defeasible logic that has been used to model the way reasons relate to requirements with the core insight of dual-role views of reasons, or views that say, roughly, that while normative reasons can play two different roles in determining all-things-considered statuses of acts, they need not always play both of them. The unified solution to the puzzles that the system supports has several advantages over related informal solutions.

While Deep Neural Networks (DNNs) can perform a wide range of tasks at human or greater-than-human level of competence, they are also notoriously opaque. This paper aims to shed light on both the specific nature of this opacity and what it would take to fully or partially remove it. We begin by drawing a clarificatory distinction between two basic dimensions of opacity of complex systems – internal and relational – and explain how various kinds of opacity invoked in recent discussions of DNNs can be understood in terms of these basic dimensions. We then discuss the prospects of removing opacity from DNNs following the methods of two subfields of research in AI – mechanistic interpretability and explainable artificial intelligence (XAI). We finish with a critical discussion of the relevance of what Sullivan (2022a) calls ‘link uncertainty’.

Thinking about misleading higher-order evidence naturally leads to a puzzle about epistemic rationality: If one’s total evidence can be radically misleading regarding itself, then two widely-accepted requirements of rationality come into conflict, suggesting that there are rational dilemmas. This paper focuses on an often misunderstood and underexplored response to this (and similar) puzzles, the so-called conflicting-ideals view. Drawing on work from defeasible logic, I propose understanding this view as a move away from the default meta-epistemological position according to which rationality requirements are strict and governed by a strong, but never explicitly stated logic, toward the more unconventional view, according to which requirements are defeasible and governed by a comparatively weak logic. When understood this way, the response is not committed to dilemmas.

We present a new approach to representation and acquisition of normative information for machine ethics. It combines an influential philosophical account of the fundamental structure of morality with argumentation theory and machine learning. According to the philosophical account, the deontic status of an action – whether it is required, forbidden, or permissible – is determined through the interaction of “normative reasons” of varying strengths or weights. We first provide a formal characterization of this account, by modeling it in(weighted) argumentation graphs. We then use it to model ethical learning: the basic idea is to use a set of cases for which deontic statuses are known to estimate the weights of normative reasons in operation in these cases, and to use these weight estimates to determine the deontic statuses of actions in new cases. The result is an approach that has the advantages of both bottom-up and top-down approaches to machine ethics: normative information is acquired through the inter-action with training data, and its meaning is clear. We also report the results of some initial experiments with the model.

While such epistemic rules as 'If you perceive that X, you ought to believe that X' and 'If you have outstanding testimony that X, you ought to believe that X' seem to be getting at important truths, it is easy to think of cases in which they come into conflict. To avoid classifying such cases as dilemmas, one can hold either that epistemic rules have built-in unless-clauses listing the circumstances under which they don't apply, or, alternatively, that epistemic rules are contributory. This paper explores both responses from a formal perspective, drawing on a simple defeasible logic framework.

According to conciliatory views on the significance of disagreement, it’s rational for you to become less confident in your take on an issue in case your epistemic peer’s take on it is different. These views are intuitively appealing, but they also face a powerful objection: in scenarios that involve disagreements over their own correctness, conciliatory views appear to self-defeat and, thereby, issue inconsistent recommendations. This paper provides a response to this objection. Drawing on the work from defeasible logics paradigm and abstract argumentation, it develops a formal model of conciliatory reasoning and explores its behavior in the troubling scenarios. The model suggests that the recommendations that conciliatory views issue in such scenarios are perfectly reasonable—even if outwardly they may look odd.

It's natural to think that the principles expressed by the statements "Promises ought to be kept" and "We ought to help those in need" are defeasible. But how are we to make sense of this defeasibility? On one proposal, moral principles have hedges or built-in unless clauses specifying the conditions under which the principle doesn't apply. On another, such principles are contributory and, thus, do not specify which actions ought to be carried out, but only what counts in favor or against them. Drawing on a defeasible logic framework, this paper sets up three models: one model for each proposal, as well as a third model capturing a mixed view on principles that combines them. It then explores the structural connections between the three models and establishes some equivalence results, suggesting that the seemingly different views captured by the models are closer than standardly thought.

Conciliatory views of disagreement say, roughly, that it’s rational for you to become less confident in your take on an issue in case you find out that an epistemic peer’s take on it is the opposite. Their intuitive appeal notwithstanding, there are well-known worries about the behavior of conciliatory views in scenarios involving higher-order disagreements, which include disagreements over these views themselves and disagreements over the peer status of alleged epistemic peers. This paper does two things. First, it explains how the core idea behind conciliatory views can be expressed in a defeasible logic framework. The result is a formal model that’s particularly useful for thinking about the behavior of conciliatory views in cases involving higher-order disagreements. And second, the paper uses this model to resolve three paradoxes associated with disagreements over epistemic peerhood.

Dialogue semantics for logic are two-player logic games between a Proponent who puts forward a logical formula φ as valid or true and an Opponent who disputes this. An advantage of the dialogical approach is that it is a uniform framework from which different logics can be obtained through only small variations of the basic rules. We introduce the composition problem for dialogue games as the problem of resolving, for a set S of rules for dialogue games, whether the set of S-dialogically valid formulas is closed under modus ponens. Solving the composition problem is fundamental for the dialogical approach to logic; despite its simplicity, it often requires an indirect solution with the help of significant logical machinery such as cut-elimination. Direct solutions to the composition problem can, however, sometimes be had. As an example, we give a set N of dialogue rules which is well-justified from the dialogical point of view, but whose set N of dialogically valid formulas is both non-trivial and non-standard. We prove that the composition problem for N can be solved directly, and introduce a tableaux system for N.

In her recent `Facts about incoherence as non-evidential epistemic reasons‘ Eva Schmidt defends the claim that not all epistemic reasons are provided by evidence. Schmidt presents three cases describing agents with incoherent beliefs and argues that, in each case, the fact that an agent’s beliefs are incoherent provides her with a non-evidential epistemic reason to suspend judgment on the issue that her beliefs are about. While I find the suggestion that facts about incoherence can play positive roles in our cognitive lives intriguing, I have three reservations about Schmidt’s view: the first concerns her conceptual framework—I think it is less neutral than it appears to be—the second concerns the view’s behavior in certain kinds of scenarios involving higher-order evidence, and the third has to do with some implausible consequences of the view. I also hint at an alternative account of the positive role of facts about incoherence.

This work explores some ways in which logics for defeasible reasoning can be applied to questions in epistemology. It's naturally thought of as developing four applications: The first is concerned with simple epistemic rules, such as "If you perceives that X, then you ought to believe that X" and "If you have outstanding testimony that X, then you ought to believe that X." Anyone who thinks that such rules have a place in our accounts of epistemic normativity must explain what happens in cases where they come into conflict —such as one where you perceive a red object and are told that it is blue. The literature has gone in two directions: The first suggests that rules have built-in unless-clauses specifying the circumstances under which they fail to apply; the second that rules do not specify what attitudes you ought to have, but only what counts in favor or against having those attitudes. I express these two different ideas in a defeasible logic framework and demonstrate that there's a clear sense in which they are equivalent. The second application uses a defeasible logic to solve an important puzzle about epistemic rationality, involving higher-order evidence, or, roughly, evidence about our capacities for evaluating evidence. My solution has some affinities with the so-called conflicting-ideals view. The third application, then, is a characterization of this view in logical terms: I suggest that it should be thought of as an unconventional metaepistemological view, according to which epistemic requirements are not exceptionless, but defeasible and governed by a comparatively weak logic. Finally, the fourth application is in the burgeoning debate about the epistemic significance of disagreement. The intuitive conciliatory views say, roughly, that you ought to become less confident in your take on some question X, if you learn that an epistemic equal disagrees with you about X. I propose to think of conciliationism as a defeasible reasoning policy, develop a mathematically precise model of it, and use it to solve one of the most pressing problems for conciliatory views: Given that there are disagreements about these views themselves, they can self-defeat and issue inconsistent recommendations.