This chapter subsumes David Marr’s levels of analysis account of explanation in cognitive science under the framework of mechanistic explanation: Answering the questions that define each one of Marr’s three levels is tantamount to describing the component parts and operations of mechanisms, as well as their organization, behavior, and environmental context. By explicating these questions and showing how they are answered in several different cognitive science research programs, this chapter resolves some of the ambiguities that remain in Marr’s account, and (...) shows that many different areas and traditions of cognitive scientific research can be unified under the mechanistic framework. (shrink)

Many of the computing systems programmed using Machine Learning are opaque: it is difficult to know why they do what they do or how they work. Explainable Artificial Intelligence aims to develop analytic techniques that render opaque computing systems transparent, but lacks a normative framework with which to evaluate these techniques’ explanatory successes. The aim of the present discussion is to develop such a framework, paying particular attention to different stakeholders’ distinct explanatory requirements. Building on an analysis of “opacity” from (...) philosophy of science, this framework is modeled after accounts of explanation in cognitive science. The framework distinguishes between the explanation-seeking questions that are likely to be asked by different stakeholders, and specifies the general ways in which these questions should be answered so as to allow these stakeholders to perform their roles in the Machine Learning ecosystem. By applying the normative framework to recently developed techniques such as input heatmapping, feature-detector visualization, and diagnostic classification, it is possible to determine whether and to what extent techniques from Explainable Artificial Intelligence can be used to render opaque computing systems transparent and, thus, whether they can be used to solve the Black Box Problem. (shrink)

Bayesian reverse-engineering is a research strategy for developing three-level explanations of behavior and cognition. Starting from a computational-level analysis of behavior and cognition as optimal probabilistic inference, Bayesian reverse-engineers apply numerous tweaks and heuristics to formulate testable hypotheses at the algorithmic and implementational levels. In so doing, they exploit recent technological advances in Bayesian artificial intelligence, machine learning, and statistics, but also consider established principles from cognitive psychology and neuroscience. Although these tweaks and heuristics are highly pragmatic in character and (...) are often deployed unsystematically, Bayesian reverse-engineering avoids several important worries that have been raised about the explanatory credentials of Bayesian cognitive science: the worry that the lower levels of analysis are being ignored altogether; the challenge that the mathematical models being developed are unfalsifiable; and the charge that the terms ‘optimal’ and ‘rational’ have lost their customary normative force. But while Bayesian reverse-engineering is therefore a viable and productive research strategy, it is also no fool-proof recipe for explanatory success. (shrink)

Willard van Orman Quine once said that he had a preference for a desert ontology. This was in an earlier day when concerns with logical structure and ontological simplicity reigned supreme. Ontological genocide was practiced upon whole classes of upper-level or ‘derivative’ entities in the name of elegance, and we were secure in the belief that one strayed irremediably into the realm of conceptual confusion and possible error the further one got from ontic fundamentalism. In those days, one paid more (...) attention to generic worries about possible errors than to actual errors derived from distancing oneself too far from the nitty-gritty details of actual theory, actual inferences from actual data, the actual conditions under which we posited and detected entities, calibrated and ‘burned in’ instruments, identified and rejected artifacts, debugged programs and procedures, explained the mechanisms behind regularities, judged correlations to be spurious, and in general, to the real complexities and richness of actual scientific practice. (shrink)

Most philosophical accounts of emergence are incompatible with reduction. Most scientists regard a system property as emergent relative to properties of its parts if it depends upon their mode of organization-a view consistent with reduction. Emergence is a failure of aggregativity, in which ``the whole is nothing more than the sum of its parts''. Aggregativity requires four conditions, giving powerful tools for analyzing modes of organization. Differently met for different decompositions of the system, and in different degrees, the structural conditions (...) can provide evaluation criteria for choosing decompositions, ``natural kinds'', and detecting functional localization fallacies, approximations, and various biases of vulgar reductionisms. This analysis of emergence and use of these conditions as heuristics is consistent with a broader reductionistic methodology. (shrink)

ion is arguably one of the most important methods in modern science in analysing and understanding complex phenomena. In his book The Philosophy of Information, Floridi (The philosophy of information. Oxford University Press, Oxford, 2011) presents the method of levels of abstraction as the main method of the Philosophy of Information. His discussion of abstraction as a method seems inspired by the formal methods and frameworks of computer science, in which abstraction is operationalised extensively in programming languages and design methodologies. (...) Is it really clear what we should understand by levels of abstraction? How should they be specified? We will argue that levels of abstraction should be augmented with annotations, in order to express semantic information for them and reconcile the method of level of abstraction (LoA’s) with other approaches. We discuss the extended method when applied e.g. to the analysis of abstract machines. This will lead to an example in which the number of LoA’s is unbounded. (shrink)

There are three main accounts of the proper function of the low mood system (LMS): the social risk theory, the disease theory, and the propitiousness theory. Adjudicating between these accounts has proven difficult, as there is little agreement in the literature about what a theory of the LMS’s proper function is supposed to explain. In this article, drawing upon influential work on the evolution of other affective systems, such as the disgust system and the fear system, I argue that a (...) theory of the proper function of the low mood system should: (i) account for the reliable, distal causes and effects of the system’s activation, and (ii) explain how having a system that performed such a function increased fitness in ancestral environments. On this basis, I show that the proper function of the low mood system is to limit resource expenditure in relatively unpropitious circumstances, exactly as hypothesised by the propitiousness theory. (shrink)

Tenenbaum and Griffiths (2001) have proposed that their Bayesian model of generalisation unifies Shepard’s (1987) and Tversky’s (1977) similarity-based explanations of two distinct patterns of generalisation behaviours by reconciling them under a single coherent task analysis. I argue that this proposal needs refinement: instead of unifying the heterogeneous notion of psychological similarity, the Bayesian approach unifies generalisation by rendering the distinct patterns of behaviours informationally relevant. I suggest that generalisation as a Bayesian inference should be seen as a complement to, (...) instead of a replacement of, similarity-based explanations. Furthermore, I show that the unificatory powers of the Bayesian model of generalisation can contribute to the selection of one of these models of psychological similarity. (shrink)

Tenenbaum and Griffiths (Behavioral and Brain Sciences 24(4):629–640, 2001) have proposed that their Bayesian model of generalisation unifies Shepard’s (Science 237(4820): 1317–1323, 1987) and Tversky’s (Psychological Review 84(4): 327–352, 1977) similarity-based explanations of two distinct patterns of generalisation behaviours by reconciling them under a single coherent task analysis. I argue that this proposal needs refinement: instead of unifying the heterogeneous notion of psychological similarity, the Bayesian approach unifies generalisation by rendering the distinct patterns of behaviours informationally relevant. I suggest that (...) generalisation as a Bayesian inference should be seen as a complement to, instead of a replacement of, similarity-based explanations. Furthermore, I show that the unificatory powers of the Bayesian model of generalisation can contribute to the selection of one of these models of psychological similarity. (shrink)

The feature-matching approach pioneered by Amos Tversky remains a groundwork for psychological models of similarity and categorization but is rarely explicitly justified considering recent advances in thinking about cognition. While psychologists often view similarity as an unproblematic foundational concept that explains generalization and conceptual thought, long-standing philosophical problems challenging this assumption suggest that similarity derives from processes of higher-level cognition, including inference and conceptual thought. This paper addresses three specific challenges to Tversky’s approach: (i) the feature-selection problem, (ii) the problem (...) of cognitive implausibility, and (iii) the problem of unprincipled tweaking. It subsequently supports key insights from Tversky’s account based on recent developments in Bayesian modeling of cognition. A novel computational view of similarity as inference is proposed that addresses each challenge by considering the contrast class as constitutive of similarity and selecting for highly informative features. In so doing, this view illustrates the ongoing promise of the feature-matching approach in explaining perception, generalization and conceptual thought by grounding them in principles of probabilistic inference. (shrink)

The debate between the defenders of explanatory unification and explanatory pluralism has been ongoing from the beginning of cognitive science and is one of the central themes of its philosophy. Does cognitive science need a grand unifying theory? Should explanatory pluralism be embraced instead? Or maybe local integrative efforts are needed? What are the advantages of explanatory unification as compared to the benefits of explanatory pluralism? These questions, among others, are addressed in this Synthese’s special issue. In the introductory paper, (...) we discuss the background of the questions, distinguishing integrative theorizing from building unified theories. On the one hand, integrative efforts involve collaboration between various disciplines, fields, approaches, or theories. These efforts could even be quite temporary, without establishing any long-term institutionalized fields or disciplines, but could also contribute to developing new interfield theories. On the other hand, unification can rely on developing complete theories of mechanisms and representations underlying all cognition, as Newell’s “unified theories of cognition”, or may appeal to grand principles, as predictive coding. Here, we also show that unification in contemporary cognitive science goes beyond reductive unity, and may involve various forms of joint efforts and division of explanatory labor. This conclusion is one of the themes present in the content of contributions constituting the special issue. (shrink)

In this paper, we address the question of what current computers are from the point of view of human-computer interaction. In the early days of computing, the Turing machine (TM) has been the cornerstone of the understanding of computers. The TM defines what can be computed and how computation can be carried out. However, in the last decades, computers have evolved and increasingly become interactive systems, reacting in real-time to external events in an ongoing loop. We argue that the TM (...) does not provide a mechanistic explanation for interactive computing. The reason is that the fundamental phenomena relevant to interactive computing are out of the scope of classical computability theory. Part of the explanatory power of the TM relies on what we propose to call an execution model. An execution model belongs to a level of abstraction where it is possible to describe both the functional architecture and the execution in mechanistic terms. An updated execution model is warranted to provide the minimal mechanistic description for interactive computation as a counterpart of what the TM could explain regarding Church-Turing computation. It would support an explanation of the ubiquitous computing devices we know - those interacting with humans, e.g., through digital interfaces. We show that such a model is not available within interactive models of computation and that relevant abstractions and concerns are available in computer engineering but need to be identified and gathered. To fill this void, we propose to reflect on the level of abstraction required to support the mechanistic description of an interactive execution and propose some preliminary requirements. (shrink)

One approach to science treats science as a cognitive accomplishment of individuals and defines a scientific community as an aggregate of individual inquirers. Another treats science as a fundamentally collective endeavor and defines a scientist as a member of a scientific community. Distributed cognition has been offered as a framework that could be used to reconcile these two approaches. Adam Toon has recently asked if the cognitive and the social can be friends at last. He answers that they probably cannot, (...) posing objections to the would-be rapprochement. We clarify both the animosity and the tonic proposed to resolve it, ultimately arguing that worries raised by Toon and others are uncompelling. (shrink)

The philosophy of information (PI) is a new area of research with its own field of investigation and methodology. This article, based on the Herbert A. Simon Lecture of Computing and Philosophy I gave at Carnegie Mellon University in 2001, analyses the eighteen principal open problems in PI. Section 1 introduces the analysis by outlining Herbert Simon's approach to PI. Section 2 discusses some methodological considerations about what counts as a good philosophical problem. The discussion centers on Hilbert's famous analysis (...) of the central problems in mathematics. The rest of the article is devoted to the eighteen problems. These are organized into five sections: problems in the analysis of the concept of information, in semantics, in the study of intelligence, in the relation between information and nature, and in the investigation of values. (shrink)

Descriptive abstraction means omission of information from descriptions of phenomena. In this paper, I introduce a distinction between vertical and horizontal descriptive abstraction. Vertical abstracts away levels of mechanism or organization, while horizontal abstracts away details within one level of organization. The distinction is implicit in parts of the literature, but it has received insufficient attention and gone mainly unnoticed. I suggest that the distinction can be used to clarify how computational descriptions are formed in some variants of the mechanistic (...) account of physical computation. Furthermore, I suggest that, if this suggestion is adopted, it can be used to resolve what I call abstraction, hierarchy, and generality problems raised against mechanistic account of physical computation. According to the abstraction problem, the mechanistic account of physical computation is conceptually confused in claiming that physical systems process computational, abstract properties. An existing solution distinguishes between descriptive and metaphysical abstraction, suggesting that the abstraction problem unnecessarily postulates metaphysically abstract entities. The solution has been criticized for leading to what I call hierarchy and generality problems: it results in two separate hierarchies, one physical and one computational, making it problematic both to account for the generality of computational descriptions and to specify how the two hierarchies are related to each other. Adopting the vertical-horizontal distinction and the view that computational descriptions are achieved by horizontal abstraction allows one to account for the generality of computational descriptions, and to form a single hierarchy in which there are no separate hierarchies in need of integration. (shrink)

Eva Hudlicka, Int. J. Mach. Conscious., 06, 29 (2014). DOI: 10.1142/S1793843014400058.

Pseudo‐mechanistic explanations in psychology and cognitive neuroscienceThis paper focuses on the level of systems/cognitive neuroscience. It argues that the great majority of explanations in psychology and cognitive neuroscience is “pseudo‐mechanistic.” On the basis of various case studies, Hommel argues that cognitive neuroscience should move beyond what he calls an “Aristotelian phase” to become a mature “Galilean” science seeking to discover actual mechanisms of cognitive phenomena.

While the research programs in early cognitive science and artificial intelligence aimed to articulate what cognition was in ideal terms, much research in contemporary computational neuroscience looks at how and why brains fail to function as they should ideally. This focus on impairment affects how we understand David Marr's hypothesized three levels of understanding. In this essay, we suggest some refinements to Marr's distinctions using a population activity model of cortico-striatal circuitry exploring impulsivity and behavioral inhibition as a case study. (...) In particular, we urge that Marr's computational level should be redefined to include a description of how systems break down. We also underscore that feed-forward processing, cognition disconnected from behavioral context, and representations do not always drive cognition in the way that Marr originally assumed. (shrink)

ABSTRACTNew trends in behavioral decision research see researchers attempting to integrate multiple-self models of behavioral economics with dual-process and dual-system theories of cognitive psych...

In this paper I discuss Searle's claim that the computational properties of a system could never cause a system to be conscious. In the first section of the paper I argue that Searle is correct that, even if a system both behaves in a way that is characteristic of conscious agents (like ourselves) and has a computational structure similar to those agents, one cannot be certain that that system is conscious. On the other hand, I suggest that Searle's intuition that (...) it is “empirically absurd” that such a system could be conscious is unfounded. In the second section I show that Searle's attempt to show that a system's computational states could not possibly cause it to be conscious is based upon an erroneous distinction between computational and physical properties. On the basis of these two arguments, I conclude that, supposing that the behavior of conscious agents can be explained in terms of their computational properties, we have good reason to suppose that a system having computational properties similar to such agents is also conscious. (shrink)

In this paper we consider two key concepts from software engineering—‘specification’ and ‘implementation’—and explore their possible applications outside software engineering to other disciplines, notably the philosophy of action, evolutionary biology, and cognitive science. Throughout, the emphasis is on the gain in conceptual clarity that can be afforded by these concepts; it is not so much a matter of new knowledge or new theories but of a reorganization of existing knowledge and theories in a way that facilitates the transfer of insights (...) across a range of related fields. (shrink)

David Marr's (1982) three‐level analysis of computational cognition argues for three distinct levels of cognitive information processing—namely, the computational, representational, and implementational levels. But Marr's levels are—and were meant to be—descriptive, rather than interactive and dynamic. For this reason, we suggest that, had Marr been writing today, he might well have gone even farther in his analysis, including the emergence of structure—in particular, explicit structure at the conceptual level—from lower levels, and the effect of explicit emergent structures on the level (...) (or levels) that gave rise to them. The message is that today's cognitive scientists need not only to understand how emergent structures—in particular, explicit emergent structures at the cognitive level—develop but also to understand how they feed back on the sub‐structures from which they emerged. (shrink)

The use of “levels of abstraction” in philosophical analysis (levelism) has recently come under attack. In this paper, I argue that a refined version of epistemological levelism should be retained as a fundamental method, called the method of levels of abstraction. After a brief introduction, in section “Some Definitions and Preliminary Examples” the nature and applicability of the epistemological method of levels of abstraction is clarified. In section “A Classic Application of the Method ofion”, the philosophical fruitfulness of the new (...) method is shown by using Kant’s classic discussion of the “antinomies of pure reason” as an example. In section “The Philosophy of the Method of Abstraction”, the method is further specified and supported by distinguishing it from three other forms of “levelism”: (i) levels of organisation; (ii) levels of explanation and (iii) conceptual schemes. In that context, the problems of relativism and antirealism are also briefly addressed. The conclusion discusses some of the work that lies ahead, two potential limitations of the method and some results that have already been obtained by applying the method to some long-standing philosophical problems. (shrink)

The notion of levels has been widely used in discussions of cognitive science, especially in discussions of the relation of connectionism to symbolic modeling of cognition. I argue that many of the notions of levels employed are problematic for this purpose, and develop an alternative notion grounded in the framework of mechanistic explanation. By considering the source of the analogies underlying both symbolic modeling and connectionist modeling, I argue that neither is likely to provide an adequate analysis of processes at (...) the level at which cognitive theories attempt to function: One is drawn from too low a level, the other from too high a level. If there is a distinctly cognitive level, then we still need to determine what are the basic organizational principles at that level. (shrink)

Prominent evolutionary psychologists have argued that our innate psychological endowment consists of numerous domainspecific cognitive resources, rather than a few domaingeneral ones. In the light of some conceptual clarification, we examine the central inprinciple arguments that evolutionary psychologists mount against domaingeneral cognition. We conclude (a) that the fundamental logic of Darwinism, as advanced within evolutionary psychology, does not entail that the innate mind consists exclusively, or even massively, of domainspecific features, and (b) that a mixed innate cognitive economy of domainspecific (...) and domaingeneral resources remains a genuine conceptual possibility. However, an examination of evolutionary psychology's 'grain problem' reveals that there is no way of establishing a principled and robust distinction between domainspecific and domaingeneral features. Nevertheless, we show that evolutionary psychologists can and do live with this grain problem without their whole enterprise being undermined. (shrink)

Recently, many philosophers have been inclined to ascribe mentality to animals on the main grounds that they possess certain complex computational abilities. In this paper I contend that this view is misleading, since it wrongly assumes that those computational abilities demand a psychological explanation. On the contrary, they can be just characterised from a computational level of explanation, which picks up a domain of computation and information processing that is common to many computing systems but is autonomous from the domain (...) of psychology. Thus, I propose that it is possible to conceive insects and other animals as mere computing agents, without having any commitment to ascribe mentality to them. I conclude by sketching a proposal about how to draw the line between mere computing and genuine mentality. (shrink)

The use of "levels of abstraction" in philosophical analysis (levellism) has recently come under attack. In this paper, we argue that a refined version of epistemological levellism should be retained as a fundamental method, which we call the method of abstraction. After a brief introduction, in section two we make clear the nature and applicability of the (epistemological) method of levels of abstraction. In section three, we show the fruitfulness of the new method by applying it to five case studies: (...) the concept of agenthood, the Turing test, the definition of emergence, quantum observation and decidable observation. In section four, we further characterise and support the method by distinguishing it from three other forms of "levellism": (i) levels of organisation; (ii) levels of explanation and (iii) conceptual schemes. In this context, we also briefly address the problems of relativism and antirealism. In the conclusion, we indicate some of the work that lies ahead, two potential limitations of the method and some results that have already been obtained by applying the method to some long-standing philosophical problems. (shrink)

At least two distinct modeling frameworks contribute to the view that mind and brain are Bayesian: Bayesian Rational Analysis (BRA) and Hierarchical Predictive Coding (HPC). What is the relative contribution of each, and how exactly do they relate? In order to answer this question, we compare the way in which these two modeling frameworks address different levels of analysis within Marr’s tripartite conception of explanation in cognitive science. Whereas BRA answers questions at the computational level only, many HPC-theorists answer questions (...) at the computational, algorithmic, and implementational levels simultaneously. Given that all three levels of analysis need to be addressed in order to explain a behavioral or cognitive phenomenon, HPC seems to deliver more complete explanations. Nevertheless, BRA is well-suited for providing a solution to the dark room problem, a major theoretical obstacle for HPC. A combination of the two approaches also combines the benefits of an embodied-externalistic approach to resolving the dark room problem with the idea of a persisting evidentiary border beyond which matters are out of cognitive reach. For this reason, the development of explanations spanning all three Marrian levels within the general Bayesian approach may require combining the BRA and HPC modeling frameworks. (shrink)

How can we explain the intentional nature of an expert’s actions, performed without immediate and conscious control, relying instead on automatic cognitive processes? How can we account for the differences and similarities with a novice’s performance of the same actions? Can a naturalist explanation of intentional expert action be in line with a philosophical concept of intentional action? Answering these and related questions in a positive sense, this dissertation develops a three-step argument. Part I considers different methods of explanations in (...) cognitive neuroscience (Bennett & Hacker’s philosophical, conceptual analysis; Marr’s three levels of explanation; Neural Correlates of Consciousness research; mechanistic explanation), defending ‘mechanistic explanation’ as a method that provides the necessary tools for integrating interdisciplinary insights into human action. Furthermore, a dynamic, explanatory mechanism allows the assessment of the impact of learning and development on expert action in a valuable way that other methods don’t. Part II continues by scrutinizing several cognitive neuroscientific theories of learning and development (neuroconstructivism; dual-processing theories; simulation theory; extended mind/cognition hypothesis), arguing for the complex interactions between different types of processing and different action representations involved in expert action performances. Moreover, according to our discussion of a particular ‘simulation theory’ these interactions can be influenced in several ways with the use of language, allowing an agent to configure a specific action representation for performance at a later stage. The results of Parts I and II are then applied in Part III to a parallel discussion of philosophical analyses of intentional action (discussing i.a. Frankfurt, Bratman, Pacherie and Ricoeur) and cognitive neuroscientific insights in it. Both approaches are found to converge in emphasizing the importance for an expert to develop stable patterns of actions that comply maximally not only with his intentions, but also with his motor expertise and with situational conditions. Consequently, his actions – automatic, or not – rely on this ‘sculpted space of actions’. (shrink)