The combination of increased availability of large amounts of fine-grained human behavioral data and advances in machine learning is presiding over a growing reliance on algorithms to address complex societal problems. Algorithmic decision-making processes might lead to more objective and thus potentially fairer decisions than those made by humans who may be influenced by greed, prejudice, fatigue, or hunger. However, algorithmic decision-making has been criticized for its potential to enhance discrimination, information and power asymmetry, and opacity. In this paper, we (...) provide an overview of available technical solutions to enhance fairness, accountability, and transparency in algorithmic decision-making. We also highlight the criticality and urgency to engage multi-disciplinary teams of researchers, practitioners, policy-makers, and citizens to co-develop, deploy, and evaluate in the real-world algorithmic decision-making processes designed to maximize fairness and transparency. In doing so, we describe the Open Algortihms project as a step towards realizing the vision of a world where data and algorithms are used as lenses and levers in support of democracy and development. (shrink)

When scientists seek further confirmation of their results, they often attempt to duplicate the results using diverse means. To the extent that they are successful in doing so, their results are said to be robust. This paper investigates the logic of such "robustness analysis" [RA]. The most important and challenging question an account of RA can answer is what sense of evidential diversity is involved in RAs. I argue that prevailing formal explications of such diversity are unsatisfactory. I propose a (...) unified, explanatory account of diversity in RAs. The resulting account is, I argue, truer to actual cases of RA in science; moreover, this account affords us a helpful, new foothold on the logic undergirding RAs. (shrink)

Scientific understanding, this paper argues, can be analyzed entirely in terms of a mental act of “grasping” and a notion of explanation. To understand why a phenomenon occurs is to grasp a correct explanation of the phenomenon. To understand a scientific theory is to be able to construct, or at least to grasp, a range of potential explanations in which that theory accounts for other phenomena. There is no route to scientific understanding, then, that does not go by way of (...) scientific explanation. (shrink)

The official model of explanation proposed by the logical empiricists, the covering law model, is subject to familiar objections. The goal of the present paper is to explore an unofficial view of explanation which logical empiricists have sometimes suggested, the view of explanation as unification. I try to show that this view can be developed so as to provide insight into major episodes in the history of science, and that it can overcome some of the most serious difficulties besetting the (...) covering law model. (shrink)

This paper argues that the medical conception of health as absence of disease is a value-free theoretical notion. Its main elements are biological function and statistical normality, in contrast to various other ideas prominent in the literature on health. Apart from universal environmental injuries, diseases are internal states that depress a functional ability below species-typical levels. Health as freedom from disease is then statistical normality of function, i.e., the ability to perform all typical physiological functions with at least typical efficiency. (...) This conception of health is as value-free as statements of biological function. The view that health is essentially value-laden, held by most writers on the topic, seems to have one of two sources: an assumption that health judgments must be practical judgments about the treatment of patients, or a commitment to "positive" health beyond the absence of disease. I suggest that the assumption is mistaken, the commitment possibly misdescribed. (shrink)

The concept of mechanism is analyzed in terms of entities and activities, organized such that they are productive of regular changes. Examples show how mechanisms work in neurobiology and molecular biology. Thinking in terms of mechanisms provides a new framework for addressing many traditional philosophical issues: causality, laws, explanation, reduction, and scientific change.

The use of black box algorithms in medicine has raised scholarly concerns due to their opaqueness and lack of trustworthiness. Concerns about potential bias, accountability and responsibility, patient autonomy and compromised trust transpire with black box algorithms. These worries connect epistemic concerns with normative issues. In this paper, we outline that black box algorithms are less problematic for epistemic reasons than many scholars seem to believe. By outlining that more transparency in algorithms is not always necessary, and by explaining that (...) computational processes are indeed methodologically opaque to humans, we argue that the reliability of algorithms provides reasons for trusting the outcomes of medical artificial intelligence. To this end, we explain how computational reliabilism, which does not require transparency and supports the reliability of algorithms, justifies the belief that results of medical AI are to be trusted. We also argue that several ethical concerns remain with black box algorithms, even when the results are trustworthy. Having justified knowledge from reliable indicators is, therefore, necessary but not sufficient for normatively justifying physicians to act. This means that deliberation about the results of reliable algorithms is required to find out what is a desirable action. Thus understood, we argue that such challenges should not dismiss the use of black box algorithms altogether but should inform the way in which these algorithms are designed and implemented. When physicians are trained to acquire the necessary skills and expertise, and collaborate with medical informatics and data scientists, black box algorithms can contribute to improving medical care. (shrink)

Organizations such as the EU High-Level Expert Group on AI and the IEEE have recently formulated ethical principles and (moral) values that should be adhered to in the design and deployment of artificial intelligence (AI). These include respect for autonomy, non-maleficence, fairness, transparency, explainability, and accountability. But how can we ensure and verify that an AI system actually respects these values? To help answer this question, I propose an account for determining when an AI system can be said to embody (...) certain values. This account understands embodied values as the result of design activities intended to embed those values in such systems. AI systems are here understood as a special kind of sociotechnical system that, like traditional sociotechnical systems, are composed of technical artifacts, human agents, and institutions but—in addition—contain artificial agents and certain technical norms that regulate interactions between artificial agents and other elements of the system. The specific challenges and opportunities of embedding values in AI systems are discussed, and some lessons for better embedding values in AI systems are drawn. (shrink)

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. (shrink)

One puzzle concerning highly idealized models is whether they explain. Some suggest they provide so-called ‘how-possibly explanations’. However, this raises an important question about the nature of how-possibly explanations, namely what distinguishes them from ‘normal’, or how-actually, explanations? I provide an account of how-possibly explanations that clarifies their nature in the context of solving the puzzle of model-based explanation. I argue that the modal notions of actuality and possibility provide the relevant dividing lines between how-possibly and how-actually explanations. Whereas how-possibly (...) explanations establish claims of possible explanations, how-actually explanations establish claims of actual ones. Models, in turn, simply provide evidence for these claims. (shrink)

This article aims to develop a new account of scientific explanation for computer simulations. To this end, two questions are answered: what is the explanatory relation for computer simulations? And what kind of epistemic gain should be expected? For several reasons tailored to the benefits and needs of computer simulations, these questions are better answered within the unificationist model of scientific explanation. Unlike previous efforts in the literature, I submit that the explanatory relation is between the simulation model and the (...) results of the simulation. I also argue that our epistemic gain goes beyond the unificationist account, encompassing a practical dimension as well. (shrink)

This article discusses the relation between features of empirical data and structures in the world. I defend the following claims. Any empirical data set exhibits all possible patterns, each with a certain noise term. The magnitude and other properties of this noise term are irrelevant to the evidential status of a pattern: all patterns exhibited in empirical data constitute evidence of structures in the world. Furthermore, distinct patterns constitute evidence of distinct structures in the world. It follows that the world (...) must be regarded as containing all possible structures. The remainder of the article is devoted to elucidating the meaning and implications of the latter claim. (shrink)

This article examines the role of experimental generalizations and physical laws in neuroscientific explanations, using Hodgkin and Huxley’s electrophysiological model from 1952 as a test case. I show that the fact that the model was partly fitted to experimental data did not affect its explanatory status, nor did the false mechanistic assumptions made by Hodgkin and Huxley. The model satisfies two important criteria of explanatory status: it contains invariant generalizations and it is modular (both in James Woodward’s sense). Further, I (...) argue that there is a sense in which the explanatory heteronomy thesis holds true for this case. †To contact the author, please write to: SNF‐Professorship for Philosophy of Science, University of Basel, Missionsstrasse 21, 4003 Basel, Switzerland; e‐mail: [email protected]. (shrink)

Comparisons of rival explanations or theories often involve vague appeals to explanatory power. In this paper, we dissect this metaphor by distinguishing between different dimensions of the goodness of an explanation: non-sensitivity, cognitive salience, precision, factual accuracy and degree of integration. These dimensions are partially independent and often come into conflict. Our main contribution is to go beyond simple stipulation or description by explicating why these factors are taken to be explanatory virtues. We accomplish this by using the contrastive-counterfactual approach (...) to explanation and the view of understanding as an inferential ability. By combining these perspectives, we show how the explanatory power of an explanation in a given dimension can be assessed by showing the range of answers it provides to what-if-things-had-been-different questions and the theoretical and pragmatic importance of these questions. Our account also explains intuitions linking explanation to unification or to exhibition of a mechanism. (shrink)

The original proof of the four-color theorem by Appel and Haken sparked a controversy when Tymoczko used it to argue that the justification provided by unsurveyable proofs carried out by computers cannot be a priori. It also created a lingering impression to the effect that such proofs depend heavily for their soundness on large amounts of computation-intensive custom-built software. Contra Tymoczko, we argue that the justification provided by certain computerized mathematical proofs is not fundamentally different from that provided by surveyable (...) proofs, and can be sensibly regarded as a priori. We also show that the aforementioned impression is mistaken because it fails to distinguish between proof search (the context of discovery) and proof checking (the context of justification). By using mechanized proof assistants capable of producing certificates that can be independently checked, it is possible to carry out complex proofs without the need to trust arbitrary custom-written code. We only need to trust one fixed, small, and simple piece of software: the proof checker. This is not only possible in principle, but is in fact becoming a viable methodology for performing complicated mathematical reasoning. This is evinced by a new proof of the four-color theorem that appeared in 2005, and which was developed and checked in its entirety by a mechanical proof system. (shrink)

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. (shrink)

In computer simulations of physical systems, the construction of models is guided, but not determined, by theory. At the same time simulations models are often constructed precisely because data are sparse. They are meant to replace experiments and observations as sources of data about the world; hence they cannot be evaluated simply by being compared to the world. So what can be the source of credibility for simulation models? I argue that the credibility of a simulation model comes not only (...) from the credentials supplied to it by the governing theory, but also from the antecedently established credentials of the model building techniques employed by the simulationists. In other words, there are certain sorts of model building techniques which are taken, in and of themselves, to be reliable. Some of these model building techniques, moreover, incorporate what are sometimes called “falsifications.” These are contrary-to-fact principles that are included in a simulation model and whose inclusion is taken to increase the reliability of the results. The example of a falsification that I consider, called artificial viscosity, is in widespread use in computational fluid dynamics. Artificial viscosity, I argue, is a principle that is successfully and reliably used across a wide domain of fluid dynamical applications, but it does not offer even an approximately “realistic” or true account of fluids. Artificial viscosity, therefore, is a counter-example to the principle that success implies truth – a principle at the foundation of scientific realism. It is an example of reliability without truth. (shrink)

We document the influence on programming language semantics of the Platonism/formalism divide in the philosophy of mathematics.

This article introduces and defends a probabilistic measure of the explanatory power that a particular explanans has over its explanandum. To this end, we propose several intuitive, formal conditions of adequacy for an account of explanatory power. Then, we show that these conditions are uniquely satisfied by one particular probabilistic function. We proceed to strengthen the case for this measure of explanatory power by proving several theorems, all of which show that this measure neatly corresponds to our explanatory intuitions. Finally, (...) we briefly describe some promising future projects inspired by our account. (shrink)

In a time when we as a society are in the process of deciding what our basic rights to health care are, it is critically important for us to have a full and complete understanding of what constitutes health. We argue for an analysis of health according to which certain states are healthy not in themselves but because they allow an individual to reach actual goals. Recognizing that the goals of an individual considered from the point of view of biology (...) and the goals of the same individual considered as an agent in the world might be different, we introduce a distinction between the health of an individual qua organism and the health of an individual qua person. We then argue that this distinction characterizes the evaluations made by patients and healthcare providers better than the widely discussed distinction between disease and illness. (shrink)