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)

In recent decades, non-representational approaches to mental phenomena and cognition have been gaining traction in cognitive science and philosophy of mind. In these alternative approach, mental representations either lose their central status or, in its most radical form, are banned completely. While there is growing agreement that non-representational accounts may succeed in explaining some cognitive capacities, there is widespread skepticism about the possibility of giving non-representational accounts of cognitive capacities such as memory, imagination or abstract thought. In this paper, I (...) will critically examine the view that there are fundamental limitations to non-representational explanations of cognition. Rather than challenging these arguments on general grounds, I will examine a set of human cognitive capacities that are generally thought to fall outside the scope of non-representational accounts, i.e. numerical cognition. After criticizing standard representational accounts of numerical cognition for their lack of explanatory power, I will argue that a non-representational approach that is inspired by radical enactivism offers the best hope for developing a genuine naturalistic explanatory account for these cognitive capacities. (shrink)

Recent accounts of emotional action intend to explain such actions without reference to goals. Nevertheless, these accounts fail to specify the difference between goals and other kinds of motivational states. I offer two remedies. First, I develop an account of goals based on Michael Smith’s arguments for the Humean theory of motivation. On this account, a goal is a unified representation that determines behavior selection criteria and satisfaction conditions for an action. This opens the possibility that mental processes could influence (...) behavior without such a unified representation and hence, without goals. Second, I develop a model of emotions and appetites on which behavior selection criteria can be decoupled from satisfaction conditions. If this model is correct, then in many cases, there is no unified representation that determines the behavior selection criteria and satisfaction conditions of emotional actions. In contrast with many traditional accounts of action, the model suggests the following: Whether or not a behavior constitutes an action does not depend on the agent’s grasp of the behavior’s aim. Rather, a behavior constitutes an action if it was organized by the agent, where an agent can organize actions without a coherent grasp of their aim. Some emotional actions are manifestations of this possibility. (shrink)

Naturalistically minded philosophers hope to identify a privileged nonsemantic relation that holds between a mental representation m and that which m represents, a relation whose privileged status underwrites the assignment of reference to m. The naturalist can accomplish this task only if she has in hand a nonsemantic criterion for individuating mental representations: it would be question-begging for the naturalist to characterize m, for the purpose of assigning content, as 'the representation with such and such content'. If we individuate mental (...) representations using the tools of dynamical systems theory, we find that a given mental representation, characterized nonsemantically, emerges in the cognitive system as the result of causal interactions between the subject and her environment. At least for the most basic of our mental representations, I argue that the dynamical systems-based approach to individuation increases the plausibility of a theory that assigns reference as a function of the subject's causal history. (shrink)

In this paper I defend the propriety of explaining the behavior of distributed connectionist networks by appeal to selected data stored therein. In particular, I argue that if there is a problem with such explanations, it is a consequence of the fact that information storage in networks is superpositional, and not because it is distributed. I then develop a ``proto-account'''' of causation for networks, based on an account of Andy Clark''s, that shows even superpositionality does not undermine information-based explanation. Finally, (...) I argue that the resulting explanations are genuinely informative and not vacuous. (shrink)

I review a widely accepted argument to the conclusion that the contents of our beliefs, desires and other mental states cannot be causally efficacious in a classical computational model of the mind. I reply that this argument rests essentially on an assumption about the nature of neural structure that we have no good scientific reason to accept. I conclude that computationalism is compatible with wide semantic causal efficacy, and suggest how the computational model might be modified to accommodate this possibility.

Structural representations are increasingly popular in philosophy of cognitive science. A key virtue they seemingly boast is that of meeting Ramsey's job description challenge. For this reason, structural representations appear tailored to play a clear representational role within cognitive architectures. Here, however, I claim that structural representations do not meet the job description challenge. This is because even our most demanding account of their functional profile is satisfied by at least some receptors, which paradigmatically fail the job description challenge. Hence, (...) the functional profile typically associated with structural representations does not identify representational posits. After a brief introduction, I present, in the second section of the paper, the job description challenge. I clarify why receptors fail to meet it and highlight why, as a result, they should not be considered representations. In the third section I introduce what I take to be the most demanding account of structural representations at our disposal, namely Gładziejewski's account. Provided the necessary background, I turn from exposition to criticism. In the first half of the fourth section, I equate the functional profile of structural representations and receptors. To do so, I show that some receptors boast, as a matter of fact, all the functional features associated with structural representations. Since receptors function merely as causal mediators, I conclude structural representations are mere causal mediators too. In the second half of the fourth section I make this conclusion intuitive with a toy example. I then conclude the paper, anticipating some objections my argument invites. (shrink)

We provide a taxonomy of the two most important debates in the philosophy of the cognitive and neural sciences. The first debate is over methodological individualism: is the object of the cognitive and neural sciences the brain, the whole animal, or the animal--environment system? The second is over explanatory style: should explanation in cognitive and neural science be reductionist-mechanistic, inter-level mechanistic, or dynamical? After setting out the debates, we discuss the ways in which they are interconnected. Finally, we make some (...) recommendations that we hope will help philosophers interested in the cognitive and neural sciences to avoid dead ends. (shrink)

Jerry Fodor and Ernest Lepore [(1992) Holism: a shopper's guide, Oxford: Blackwell; (1996) in R. McCauley (Ed.) The Churchlands and their critics , Cambridge: Blackwell] have launched a powerful attack against Paul Churchland's connectionist theory of semantics--also known as state space semantics. In one part of their attack, Fodor and Lepore argue that the architectural and functional idiosyncrasies of connectionist networks preclude us from articulating a notion of conceptual similarity applicable to state space semantics. Aarre Laakso and Gary Cottrell [(1998) (...) in M. A. Gernsbacher & S. Derry (Eds) Proceedings of the 20th Annual Conference of the Cognitive Science Society, Mahway, NJ: Erlbaum; Philosophical Psychology ] 13, 47-76 have recently run a number of simulations on simple feedforward networks and applied a mathematical technique for measuring conceptual similarity in the representational spaces of those networks. Laakso and Cottrell contend that their results decisively refute Fodor and Lepore's criticisms. Paul Churchland [(1998) Journal of Philosophy, 95, 5-32 ] goes further. He uses Laakso and Cottrell's neurosimulations to argue that connectionism does furnish us with all we need to construct a robust theory of semantics and a robust theory of translation. In this paper I shall argue that whereas Laakso and Cottrell's neurocomputational results may provide us with a rebuttal of Fodor and Lepore's argument, Churchland's conclusion is far too optimistic. In particular, I shall try to show that connectionist modelling does not provide any objective criterion for achieving a one-to-one accurate translational mapping across networks. (shrink)

In agreement with some of the ideas expressed by Perruchet & Vinter (P&V), we believe that some phenomena hitherto attributed to processing may in fact reflect a fundamental distinction between direct and reflexive forms of consciousness. This dichotomy, developed by the phenomenological tradition, is substantiated by examples coming from experimental psychology and lesion neuropsychology.

This paper is concerned with representational explanations of how one experiences and acts with one’s body as an integrated whole. On the standard view, accounts of bodily experience and action must posit a corresponding representational structure: a representation of the body as an integrated whole. The aim of this paper is to show why we should instead favour the minimal view: given the nature of the body, and representation of its parts, accounts of the structure of bodily experience and action (...) need not appeal to a representation of the body as an integrated whole. The argument proceeds by distinguishing two kinds of explanatory roles for representations: standing-in for absent targets and structuring ambiguous sensory information concerning a target. Representations of body-parts are suited to fulfil both kinds of explanatory role, whereas a representation of the body as an integrated whole is only suited to fulfil the latter, as a means of coordinating representations of body-parts. It is then argued that the structure of the body can itself serve as a means of coordinating body-part representations, rendering representation of the body as an integrated whole explanatorily superfluous. (shrink)

Phenomenal intentionality is supposed to be a kind of directedness of the mind onto the world that is grounded in the conscious feel of mental life. This book of new essays explores a number of issues raised by the notion of phenomenal intentionality.

Distributed Cognition and Integrational Linguistics have much in common. Both approaches see communicative activity and intelligent behaviour in general as strongly con- text-dependent and action-oriented, and brains as permeated by history. But there is some ten- sion between the two frameworks on three important issues. The majority of theorists of distributed cognition want to maintain some notions of mental representation and computa- tion, and to seek generalizations and patterns in the various ways in which creatures like us couple with technologies, (...) media, and other agents; many also want to offer explanations at subpersonal levels which may undercut the autonomy of personal-level accounts. In contrast, dominant views in integrational linguistics reject all invocation of representation, resist the explanatory search for similarity across contexts and moments, and see linguistics as a lay dis- cipline which should not offer explanations in terms alien to ordinary agents. On each of these issues, I argue that integrationists could move closer to the distributed cognition frame- work without losing the most important aspects of their view: integrationist criticisms of mainstream or classical theories can be respected while alliances with revised cognitivist views about representation, context, and explanation are developed. Ó 2004 Elsevier Ltd. All rights reserved. (shrink)

David Hume's famous riddle of induction implies a second problem related to the question of whether the laws and principles of nature might change in the course of time. Claims have been made that modern developments in physics and astrophysics corroborate the translational invariance of the laws of physics in time. However, the appearance of a new general principle of nature, which might not be derivable from the known laws of physics, or that might actually be a non-physical one (this (...) means completely independent of physical science) supports the notion that the course of nature can change in time. Here it is argued that natural selection satisfies the criteria that identify a general principle of nature which so far, appears to be non-derivable from the known laws of physics and therefore, it is likely that it arose in the course of time, thus leaving open again the quest for a true solution to Hume's second problem. (shrink)