We propose a formal framework for interpretable machine learning. Combining elements from statistical learning, causal interventionism, and decision theory, we design an idealised explanation game in which players collaborate to find the best explanation for a given algorithmic prediction. Through an iterative procedure of questions and answers, the players establish a three-dimensional Pareto frontier that describes the optimal trade-offs between explanatory accuracy, simplicity, and relevance. Multiple rounds are played at different levels of abstraction, allowing the players to explore overlapping causal (...) patterns of variable granularity and scope. We characterise the conditions under which such a game is almost surely guaranteed to converge on a optimal explanation surface in polynomial time, and highlight obstacles that will tend to prevent the players from advancing beyond certain explanatory thresholds. The game serves a descriptive and a normative function, establishing a conceptual space in which to analyse and compare existing proposals, as well as design new and improved solutions. (shrink)

Machine learning algorithms may radically improve our ability to diagnose and treat disease. For moral, legal, and scientific reasons, it is essential that doctors and patients be able to understand and explain the predictions of these models. Scalable, customisable, and ethical solutions can be achieved by working together with relevant stakeholders, including patients, data scientists, and policy makers.

We propose a formal framework for interpretable machine learning. Combining elements from statistical learning, causal interventionism, and decision theory, we design an idealisedexplanation gamein which players collaborate to find the best explanation(s) for a given algorithmic prediction. Through an iterative procedure of questions and answers, the players establish a three-dimensional Pareto frontier that describes the optimal trade-offs between explanatory accuracy, simplicity, and relevance. Multiple rounds are played at different levels of abstraction, allowing the players to explore overlapping causal patterns of (...) variable granularity and scope. We characterise the conditions under which such a game is almost surely guaranteed to converge on a (conditionally) optimal explanation surface in polynomial time, and highlight obstacles that will tend to prevent the players from advancing beyond certain explanatory thresholds. The game serves a descriptive and a normative function, establishing a conceptual space in which to analyse and compare existing proposals, as well as design new and improved solutions. (shrink)

As machine learning has gradually entered into ever more sectors of public and private life, there has been a growing demand for algorithmic explainability. How can we make the predictions of complex statistical models more intelligible to end users? A subdiscipline of computer science known as interpretable machine learning (IML) has emerged to address this urgent question. Numerous influential methods have been proposed, from local linear approximations to rule lists and counterfactuals. In this article, I highlight three conceptual challenges that (...) are largely overlooked by authors in this area. I argue that the vast majority of IML algorithms are plagued by (1) ambiguity with respect to their true target; (2) a disregard for error rates and severe testing; and (3) an emphasis on product over process. Each point is developed at length, drawing on relevant debates in epistemology and philosophy of science. Examples and counterexamples from IML are considered, demonstrating how failure to acknowledge these problems can result in counterintuitive and potentially misleading explanations. Without greater care for the conceptual foundations of IML, future work in this area is doomed to repeat the same mistakes. (shrink)

Artificial intelligence has historically been conceptualized in anthropomorphic terms. Some algorithms deploy biomimetic designs in a deliberate attempt to effect a sort of digital isomorphism of the human brain. Others leverage more general learning strategies that happen to coincide with popular theories of cognitive science and social epistemology. In this paper, I challenge the anthropomorphic credentials of the neural network algorithm, whose similarities to human cognition I argue are vastly overstated and narrowly construed. I submit that three alternative supervised learning (...) methods—namely lasso penalties, bagging, and boosting—offer subtler, more interesting analogies to human reasoning as both an individual and a social phenomenon. Despite the temptation to fall back on anthropomorphic tropes when discussing AI, however, I conclude that such rhetoric is at best misleading and at worst downright dangerous. The impulse to humanize algorithms is an obstacle to properly conceptualizing the ethical challenges posed by emerging technologies. (shrink)

We propose a formal framework for interpretable machine learning. Combining elements from statistical learning, causal interventionism, and decision theory, we design an idealised explanation game in which players collaborate to find the best explanation for a given algorithmic prediction. Through an iterative procedure of questions and answers, the players establish a three-dimensional Pareto frontier that describes the optimal trade-offs between explanatory accuracy, simplicity, and relevance. Multiple rounds are played at different levels of abstraction, allowing the players to explore overlapping causal (...) patterns of variable granularity and scope. We characterise the conditions under which such a game is almost surely guaranteed to converge on a optimal explanation surface in polynomial time, and highlight obstacles that will tend to prevent the players from advancing beyond certain explanatory thresholds. The game serves a descriptive and a normative function, establishing a conceptual space in which to analyse and compare existing proposals, as well as design new and improved solutions. (shrink)

Unsupervised learning algorithms are widely used for many important statistical tasks with numerous applications in science and industry. Yet despite their prevalence, they have attracted remarkably little philosophical scrutiny to date. This stands in stark contrast to supervised and reinforcement learning algorithms, which have been widely studied and critically evaluated, often with an emphasis on ethical concerns. In this article, I analyze three canonical unsupervised learning problems: clustering, abstraction, and generative modeling. I argue that these methods raise unique epistemological and (...) ontological questions, providing data-driven tools for discovering natural kinds and distinguishing essence from contingency. This analysis goes some way toward filling the lacuna in contemporary philosophical discourse on unsupervised learning, as well as bringing conceptual unity to a heterogeneous field more often described by what it isnot(i.e., supervised or reinforcement learning) than by what itis. I submit that unsupervised learning is not just a legitimate subject of philosophical inquiry but perhaps the most fundamental branch of all AI. However, an uncritical overreliance on unsupervised methods poses major epistemic and ethical risks. I conclude by advocating for a pragmatic, error-statistical approach that embraces the opportunities and mitigates the challenges posed by this powerful class of algorithms. (shrink)

Online controlled experiments, also known as A/B tests, have become ubiquitous. While many practical challenges in running experiments at scale have been thoroughly discussed, the ethical dimension of A/B testing has been neglected. This article fills this gap in the literature by introducing a new, soft ethics and governance framework that explicitly recognizes how the rise of an experimentation culture in industry settings brings not only unprecedented opportunities to businesses but also significant responsibilities. More precisely, the article (a) introduces a (...) set of principles to encourage ethical and responsible experimentation to protect users, customers, and society; (b) argues that ensuring compliance with the proposed principles is a complex challenge unlikely to be addressed by resorting to a one-solution response; (c) discusses the relevance and effectiveness of several mechanisms and policies in educating, governing, and incentivizing companies conducting online controlled experiments; and (d) offers a list of prompting questions specifically designed to help and empower practitioners by stimulating specific ethical deliberations and facilitating coordination among different groups of stakeholders. (shrink)

Necessity and sufficiency are the building blocks of all successful explanations. Yet despite their importance, these notions have been conceptually underdeveloped and inconsistently applied in explainable artificial intelligence, a fast-growing research area that is so far lacking in firm theoretical foundations. In this article, an expanded version of a paper originally presented at the 37th Conference on Uncertainty in Artificial Intelligence, we attempt to fill this gap. Building on work in logic, probability, and causality, we establish the central role of (...) necessity and sufficiency in XAI, unifying seemingly disparate methods in a single formal framework. We propose a novel formulation of these concepts, and demonstrate its advantages over leading alternatives. We present a sound and complete algorithm for computing explanatory factors with respect to a given context and set of agentive preferences, allowing users to identify necessary and sufficient conditions for desired outcomes at minimal cost. Experiments on real and simulated data confirm our method’s competitive performance against state of the art XAI tools on a diverse array of tasks. (shrink)

We explore the implications of a hierarchical structure, consisting of the higher order dimensions of nonspecific Positive Activation and Negative Activation and multiple specific negative affects and positive affects at the lower level. Emotional blends of the same valence are an essential part of this structure and form the basis of the higher order Negative and Positive Activation dimensions. Mixed cross-valence emotions are not central to this hierarchical scheme but are compatible with it. We examine the frequency of pure emotional (...) states, same-valence emotional blends, and cross-valence mixed emotions in a large momentary mood sample. (shrink)

(No abstract is available for this citation).

Recent years have seen a surge in online collaboration between experts and amateurs on scientific research. In this article, we analyse the epistemological implications of these crowdsourced projects, with a focus on Zooniverse, the world’s largest citizen science web portal. We use quantitative methods to evaluate the platform’s success in producing large volumes of observation statements and high impact scientific discoveries relative to more conventional means of data processing. Through empirical evidence, Bayesian reasoning, and conceptual analysis, we show how information (...) and communication technologies enhance the reliability, scalability, and connectivity of crowdsourced e-research, giving online citizen science projects powerful epistemic advantages over more traditional modes of scientific investigation. These results highlight the essential role played by technologically mediated social interaction in contemporary knowledge production. We conclude by calling for an explicitly sociotechnical turn in the philosophy of science that combines insights from statistics and logic to analyse the latest developments in scientific research. (shrink)

Discovering high-level causal relations from low-level data is an important and challenging problem that comes up frequently in the natural and social sciences. In a series of papers, Chalupka etal. (2015, 2016a, 2016b, 2017) develop a procedure forcausal feature learning (CFL) in an effortto automate this task. We argue that CFL does not recommend coarsening in cases where pragmatic considerations rule in favor of it, and recommends coarsening in cases where pragmatic considerations rule against it. We propose a new technique, (...) pragmatic causal feature learning (PCFL), which extends the original CFL algorithm in useful and intuitive ways. We show that PCFL has the same attractive measure-theoretic properties as the original CFL algorithm. We compare the performance of both methods through theoretical analysis and experiments. (shrink)

Necessity and sufficiency are the building blocks of all successful explanations. Yet despite their importance, these notions have been conceptually underdeveloped and inconsistently applied in explainable artificial intelligence (XAI), a fast-growing research area that is so far lacking in firm theoretical foundations. Building on work in logic, probability, and causality, we establish the central role of necessity and sufficiency in XAI, unifying seemingly disparate methods in a single formal framework. We provide a sound and complete algorithm for computing explanatory factors (...) with respect to a given context, and demonstrate its flexibility and competitive performance against state of the art alternatives on various tasks. (shrink)

According to Lewis's modal realism, all ways the world could be are represented by possible worlds, and all possible worlds represent some way the world could be. That there are just the right possible worlds to represent all and only the ways the world could be is to be guaranteed by the principle of recombination. Lewis sketches the principle (put roughly: anything can co-exist with anything else), but does not spell out a precise version that generates just the right possibilities. (...) David Efird and Tom Stoneham have offered a principle that aims to do just that.In this paper, we argue that Efird's and Stoneham's principle of recombination is not successful – it fails to generate the right possibilities – but we also suggest ways that their account might be improved to solve the problem we raise. We also argue against Efird's and Stoneham's claim that the correct principle of recombination demonstrates the possibility of nothing concrete – it is true that their principle of recombination has models consistent with the existence of an empty world, but we only get the possibility of nothing if mereologically null individuals are possible. The Lewisian should only think mereologically null individuals are possible if he or she has some independent reason for believing in the possibility of an empty world, so the principle of recombination provides no new evidence for that possibility. We draw some morals from this for the correct way to formulate the principle of recombination. (shrink)

The modern abundance and prominence of data have led to the development of “data science” as a new field of enquiry, along with a body of epistemological reflections upon its foundations, methods, and consequences. This article provides a systematic analysis and critical review of significant open problems and debates in the epistemology of data science. We propose a partition of the epistemology of data science into the following five domains: (i) the constitution of data science; (ii) the kind of enquiry (...) that it identifies; (iii) the kinds of knowledge that data science generates; (iv) the nature and epistemological significance of “black box” problems; and (v) the relationship between data science and the philosophy of science more generally. (shrink)

Two self-paced reading experiments investigated several factors that influence the comprehension complexity of singly-embedded relative clauses (RCs) in English. Three factors were manipulated in Experiment 1, resulting in three main effects. First, object-extracted RCs were read more slowly than subject-extracted RCs, replicating previous work. Second, RCs that were embedded within the sentential complement of a noun were read more slowly than comparable RCs that were not embedded in this way. Third, and most interestingly, object-modifying RCs were read more slowly than (...) subject-modifying relative clauses. This result contradicts one of the central tenets of complexity research: that nested sentences are harder to understand than their right-branching equivalents (e.g., Miller and Chomsky 1963). It is hypothesized that this result followed from a combination of two information-flow factors: (1) background information is usually presented early in a sentence; and (2) restrictive RCs—the form of the RCs in Experiment 1—usually convey background information. Experiment 2 tested this hypothesis by comparing restrictive and non-restrictive RCs— which generally provide new information—in both subject- and objectmodifying positions. The results of the experiment were as predicted by the information-flow account: Only restrictive RCs were read more slowly when modifying objects. It is concluded that both resource and information-flow factors need to be considered in explaining RC complexity effects. (shrink)

Although the presence or absence of a pitch accent clearly can play an important role in signaling the discourse and information structure of an utterance, whether the form of an accent determines the type of information it conveys is more controversial. We used an eye‐tracking paradigm to investigate whether H*, which has been argued to signal new information, evokes different eye fixations than L+H*, which has been argued to signal the presence of contrast. Our results demonstrate that although listeners interpret (...) these accents differently, their interpretive domains overlap. L+H* creates a strong bias toward contrast referents whereas H* is compatible with both new and contrast referents. (shrink)

On the whole, the US Algorithmic Accountability Act of 2022 (US AAA) is a pragmatic approach to balancing the benefits and risks of automated decision systems. Yet there is still room for improvement. This commentary highlights how the US AAA can both inform and learn from the European Artificial Intelligence Act (EU AIA).

In a previous article, we projected the future accumulation of profiles belonging to deceased users on Facebook. We concluded that a minimum of 1.4 billion users will pass away before 2100 if Facebook ceases to attract new users as of 2018. If the network continues expanding at current rates, on the other hand, this number will exceed 4.9 billion. Although these findings provided an important first step, one network alone remains insufficient to establish a quantitative foundation for further macro-level analysis (...) of the phenomenon of online death. Facebook is but one social media platform among many, and hardly the most representative in terms of their policy on deceased users. In this study, we use the same methodology to develop a complementary analysis of projected mortality on Instagram. Our models indicate that somewhere between 767 million and 4.2 billion Instagram users will die between 2019 and 2100, depending on the network’s future growth rate. Although the number of deceased Instagram profiles will likely be fewer than those on Facebook, we argue that they are nonetheless part of a shared digital cultural heritage, and should hence be curated with careful consideration. (shrink)

The modern abundance and prominence of data has led to the development of “data science” as a new field of enquiry, along with a body of epistemological reflections upon its foundations, methods, and consequences. This article provides a systematic analysis and critical review of significant open problems and debates in the epistemology of data science. We propose a partition of the epistemology of data science into the following five domains: (i) the constitution of data science; (ii) the kind of enquiry (...) that it identifies; (iii) the kinds of knowledge that data science generates; (iv) the nature and epistemological significance of “black box” problems; and (v) the relationship between data science and the philosophy of science more generally. (shrink)

Organisations that design and deploy artificial intelligence (AI) systems increasingly commit themselves to high-level, ethical principles. However, there still exists a gap between principles and practices in AI ethics. One major obstacle organisations face when attempting to operationalise AI Ethics is the lack of a well-defined material scope. Put differently, the question to which systems and processes AI ethics principles ought to apply remains unanswered. Of course, there exists no universally accepted definition of AI, and different systems pose different ethical (...) challenges. Nevertheless, pragmatic problem-solving demands that things should be sorted so that their grouping will promote successful actions for some specific end. In this article, we review and compare previous attempts to classify AI systems for the purpose of implementing AI governance in practice. We find that attempts to classify AI systems proposed in previous literature use one of three mental models: _the Switch_, i.e., a binary approach according to which systems either are or are not considered AI systems depending on their characteristics; _the Ladder_, i.e., a risk-based approach that classifies systems according to the ethical risks they pose; and _the Matrix_, i.e., a multi-dimensional classification of systems that take various aspects into account, such as context, input data, and decision-model. Each of these models for classifying AI systems comes with its own set of strengths and weaknesses. By conceptualising different ways of classifying AI systems into simple mental models, we hope to provide organisations that design, deploy, or regulate AI systems with the vocabulary needed to demarcate the material scope of their AI governance frameworks. (shrink)

(No abstract is available for this citation).

1. The problemLewis identifies the monadic property being F with the set of all actual and possible Fs; the dyadic relation R is identified with the set of actual and possible pairs of things that are related by R; and so on . 1 Egan has argued that the fact that some properties have some of their properties contingently leads to trouble: " Let @ be the actual world, in which being green is [someone's] favourite property, and let w be (...) a world in which being green is [nobody's] favourite property. Since being green is somebody's favourite property in @, it must be a member of … being somebody's favourite property. Since being green is not anybody's favourite property in w, it must not be a member of being somebody's favourite property. Contradiction. " Egan suggests identifying properties with functions from worlds to extensions. However, as Lewis and others have complained, this has the drawback of turning monadic properties into relations . 2 In this article, we shall show that the Lewisian has the resources to block the above argument while simultaneously respecting the adicity of the relevant property and the contingency of the predication without changing his theory of properties. However, our analysis reveals a hidden ambiguity in the Lewisian conception of properties, and on either disambiguation the resulting account of properties seems inadequate.2. On ‘being someone's favourite property’The phrase ‘ being someone's favourite property’ is ambiguous. Given Lewis's ontology of possibilia, where the phrase picks out a monadic property it is not one that is possessed contingently; where it picks out a property that is possessed contingently it is not one that is monadic. But the argument works only if being someone's favourite property is both monadic and possessed contingently.Why should …. (shrink)

The ICD-11 personality disorder model is the first fully dimensional assessment of personality pathology. It consists of a personality disorder dysfunction-severity dimension, which encompasses both self- and interpersonal dysfunction, and six optional qualifiers for five prominent personality traits—Negative Affectivity, Detachment, Dissociality, Disinhibition, and Anankastia —plus a borderline pattern that is defined by the criteria of DSM-IV borderline PD. This article reports on the development of a new self-report measure to assess self- and interpersonal dysfunction and the five trait qualifiers. It (...) is the first comprehensive measure of the ICD-11 PD model in that it is the only one to include both PD dysfunction-severity as well as trait scales and because it is based on the Clinical Description and Diagnostic Guidelines, which are more detailed than the “statistical” model description that is currently on the ICD-11 website. The authors wrote 992 items and then reduced the pool to 300 items by eliminating redundancy and selecting the consensus best few items for each subconstruct. Data were collected using an online sample of 383 Prolific workers. Using exploratory factor analysis, seven domain scales were developed, each of which contained two to four scales assessing components of the domain. These preliminary scales’ psychometrics were excellent, as were the domains’ and their components’ convergent and discriminant validity, with a few generally minor exceptions. Structural analyses at the component level revealed a three-factor structure consisting of two moderately correlated Internalizing factors, one centered on Self Dysfunction with two NA components and a DSN component and the other on Interpersonal Dysfunction with DET and ANK components; as well as an Externalizing factor with DSL and a DSN component that was uncorrelated with the other two factors. Two aspects of the results in particular are striking: ANK was not the opposite end of a DSN dimension, but rather contributed to an Internalizing Interpersonal Dysfunction dimension and DSN had both an Internalizing and an Externalizing component. Implications of the findings and study limitations are discussed. (shrink)

This book explores a wide range of topics in digital ethics. It features 11 chapters that analyze the opportunities and the ethical challenges posed by digital innovation, delineate new approaches to solve them, and offer concrete guidance to harness the potential for good of digital technologies. The contributors are all members of the Digital Ethics Lab, a research environment that draws on a wide range of academic traditions. The chapters highlight the inherently multidisciplinary nature of the subject, which cannot be (...) separated from the epistemological foundations of the technologies themselves or the political implications of the requisite reforms. Coverage illustrates the importance of expert knowledge in the project of designing new reforms and political systems for the digital age. The contributions also show how this task requires a deep self-understanding of who we are as individuals and as a species. The questions raised here have ancient -- perhaps even timeless -- roots. The phenomena they address may be new. But, the contributors examine the fundamental concepts that undergird them: good and evil, justice and truth. Indeed, every epoch has its great challenges. The role of philosophy must be to redefine the meaning of these concepts in light of the particular challenges it faces. This is true also for the digital age. This book takes an important step towards redefining and re-implementing fundamental ethical concepts to this new era. (shrink)

Syntactic priming in language production is the increased likelihood of using a recently encountered syntactic structure. In this paper, we examine two theories of why speakers can be primed: error‐driven learning accounts (Bock, Dell, Chang, & Onishi, 2007; Chang, Dell, & Bock, 2006) and activation‐based accounts (Pickering & Branigan, 1999; Reitter, Keller, & Moore, 2011). Both theories predict that speakers should be primed by the syntactic choices of others, but only activation‐based accounts predict that speakers should be able to prime (...) themselves. Here we test whether speakers can be primed by their own productions in three behavioral experiments and find evidence of structural persistence following both comprehension and speakers’ own productions. We also find that comprehension‐based priming effects are larger for rarer syntactic structures than for more common ones, which is most consistent with error‐driven accounts. Because neither error‐driven accounts nor activation‐based accounts fully explain the data, we propose a hybrid model. (shrink)

Rhythmic structure in speech is characterized by sequences of stressed and unstressed syllables. A large body of literature suggests that speakers of English attempt to achieve rhythmic harmony by evenly distributing stressed syllables throughout prosodic phrases. The question remains as to how speakers plan metrical structure during speech production and whether it is planned independently of phonemes. To examine this, we designed a tongue twister task consisting of disyllabic word pairs with overlapping phonological segments and either matching or non‐matching metrical (...) structure. Results showed that speakers had more difficulty producing metrically regular word pairs, compared to irregular pairs; that is, word pairs with irregular meter had faster productions and fewer speech errors in this production task. This finding of metrical regularity inhibiting production is inconsistent with an abstract metrical structure that is planned independently of phonemes at the point of phonological encoding. (shrink)