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)

Theories concerning the structure, or format, of mental representation should (1) be formulated in mechanistic, rather than metaphorical terms; (2) do justice to several philosophical intuitions about mental representation; and (3) explain the human capacity to predict the consequences of worldly alterations (i.e., to think before we act). The hypothesis that thinking involves the application of syntax-sensitive inference rules to syntactically structured mental representations has been said to satisfy all three conditions. An alternative hypothesis is that thinking requires the construction (...) and manipulation of the cognitive equivalent of scale models. A reading of this hypothesis is provided that satisfies condition (1) and which, even though it may not fully satisfy condition (2), turns out (in light of the frame problem) to be the only known way to satisfy condition (3). (shrink)

Figure 1: A pr ototyp ical exa mple of a three-layer feed forward network, used by Plunkett and M archm an (1 991 ) to simulate learning the past-tense of En glish verbs. The inpu t units encode representations of the three phonemes of the present tense of the artificial words used in this simulation. Th e netwo rk is trained to produce a representation of the phonemes employed in the past tense form and the suffix (/d/, /ed/, or /t/) (...) used on regular verbs. To run the network, each input unit is assigned an activation value o f 0 or 1 , dep ending on whethe r the featu re is present or not. Eac h input unit is conne cted to each of the 30 hidden units by a we ighted conn ection and p rovid es an inp ut to each hidden unit equal to the product of the input unit's activation and the weight. Each hidd en unit's activation is then determined by summing ov er the va lues co ming fro m each inp ut unit to deter mine a netinput, and then applying a non-linear function (e.g., the logistic function 1/(1+e^netinput). Th is whole proced ure is. (shrink)

This introduction to a special issue of Topoi introduces and summarises the relationship between three main varieties of 'enactivist' theorising about the mind: 'autopoietic', 'sensorimotor', and 'radical' enactivism. It includes a brief discussion of the philosophical and cognitive scientific precursors to enactivist theories, and the relationship of enactivism to other trends in embodied cognitive science and philosophy of mind.

According to the thesis of semantic underdetermination, most sentences of a natural language lack a definite semantic interpretation. This thesis supports an argument against the use of natural language as an instrument of thought, based on the premise that cognition requires a semantically precise and compositional instrument. In this paper we examine several ways to construe this argument, as well as possible ways out for the cognitive view of natural language in the introspectivist version defended by Carruthers. Finally, we sketch (...) a view of the role of language in thought as a specialized tool, showing how it avoids the consequences of semantic underdetermination. (shrink)

The social brain hypothesis implies that humans and other primates evolved “modules“ for representing social knowledge. Alternatively, no such cognitive specializations are needed because social knowledge is already present in the world — we can simply monitor the dynamics of social interactions. Given the latter idea, what mechanism could account for coalition formation? We propose that statistical learning can provide a mechanism for fast and implicit learning of social signals. Using human participants, we compared learning of social signals with arbitrary (...) signals. We found that learning of social signals was no better than learning of arbitrary signals. While coupling faces and voices led to parallel learning, the same was true for arbitrary shapes and sounds. However, coupling versus uncoupling social signals with arbitrary signals revealed that faces and voices are treated with perceptual priority. Overall, our data suggest that statistical learning is a viable domain-general mechanism for learning social group structure. Keywords: social brain; embodied cognition; distributed cognition; situated cognition; multisensory; audiovisual speech; crossmodal; multimodal. (shrink)

The historical debate on representation in cognitive science and neuroscience construes representations as theoretical posits and discusses the degree to which we have reason to posit them. We reject the premise of that debate. We argue that experimental neuroscientists routinely observe and manipulate neural representations in their laboratory. Therefore, neural representations are as real as neurons, action potentials, or any other well-established entities in our ontology.

On the extended mind hypothesis (EM), many of our cognitive states and processes are hybrids, unevenly distributed across biological and nonbiological realms. In certain circumstances, things - artifacts, media, or technologies - can have a cognitive life, with histories often as idiosyncratic as those of the embodied brains with which they couple. The realm of the mental can spread across the physical, social, and cultural environments as well as bodies and brains. My independent aims in this chapter are: first, to (...) describe two compatible but distinct movements or "waves" within the EM literature, arguing for the priority of the second wave (and gesturing briefly toward a third); and, second, to defend and illustrate the interdisciplinary implications of EM as best understood, specifically for historical disciplines, by sketching two case studies. (shrink)

Multiple explanatory frameworks may be required to provide an adequate account of human cognition. This paper embeds the classical account within a neural network framework, exploring the encoding of syntacticallystructured objects over the synchronicdiachronic characteristics of networks. Synchronic structure is defined in terms of temporal binding and the superposition of states. To accommodate asymmetric relations, synchronic structure is subject to the type uniqueness constraint. The nature of synchronic structure is shown to underlie Xbar theory that characterizes the phrasal structure of (...) human languages. The derivation differentiates core Xbar properties from languagespecific properties. (shrink)

This paper advocates explicitness about the type of entity to be considered as content- bearing in connectionist systems; it makes a positive proposal about how vehicles of content should be individuated; and it deploys that proposal to argue in favour of representation in connectionist systems. The proposal is that the vehicles of content in some connectionist systems are clusters in the state space of a hidden layer. Attributing content to such vehicles is required to vindicate the standard explanation for some (...) classificatory networks’ ability to generalise to novel samples their correct classification of the samples on which they were trained. (shrink)

Despite a wealth of data we still have no clear idea what color experiences represent. In fact, color experiences vary with so many factors that it has been claimed that they do not represent anything at all. The primary challenge for any representational account of color experience is to accommodate the various psychophysical results that demonstrate that color appearance depends not only on the spectral nature of the target but also on the spectral, spatial and figural nature of the surround. (...) A number of theorists have proposed that this dependence is an aspect of the visual system's constancy mechanism. However this does not in and of itself tell us what, if anything, is represented in color experience. Ultimately the answer to this question will be informed by one's theory of representational content. I will argue that adopting a molecular scheme of representation enables the development of an account of the represented object of color experience that can do justice to the psychophysical data. (shrink)

Embedded and embodied approaches to cognition urge that (1) complicated internal representations may be avoided by letting features of the environment drive behavior, and (2) environmental structures can play an enabling role in cognition, allowing prior cognitive processes to solve novel tasks. Such approaches are thus in a natural position to oppose the ‘thesis of linguistic structuring’: The claim that the ability to use language results in a wholesale recapitulation of linguistic structure in onboard mental representation. Prominent examples of researchers (...) adopting this critical stance include Andy Clark, Michael Wheeler, and Mark Rowlands. But is such opposition warranted? Since each of these authors advocate accounts of mental representation that are broadly connectionist, I survey research on formal language computation in artificial neural networks, and argue that results indicate a strong form of the linguistic structuring thesis is true: Internal representational systems recapitulate significant linguistic structure, even on a connectionist account of mental representation. I conclude by sketching how my conclusion can nonetheless be viewed as consistent with and complimentary to an embedded/embodied account of the role of linguistic structure in cognition. (shrink)

Fodor has argued that observation is theory neutral, since the perceptual systems are modular, that is, they are domain‐specific, encapsulated, mandatory, fast, hard‐wired in the organism, and have a fixed neural architecture. Churchland attacks the theoretical neutrality of observation on the grounds that (a) the abundant top‐down pathways in the brain suggest the cognitive penetration of perception and (b) perceptual learning can change in the wiring of the perceptual systems. In this paper I introduce a distinction between sensation, perception, and (...) observation and I argue that although Churchland is right that observation involves top‐down processes, there is also a substantial amount of information in perception which is theory‐neutral. I argue that perceptual learning does not threaten the cognitive impenetrability of perception, and that the neuropsychological research does not provide evidence in favor of the top‐down character of perception. Finally, I discuss the possibility of an off‐line cognitive penetrability of perception. (shrink)

When cognitive scientists apply computational theory to the problem of phenomenal consciousness, as many of them have been doing recently, there are two fundamentally distinct approaches available. Either consciousness is to be explained in terms of the nature of the representational vehicles the brain deploys; or it is to be explained in terms of the computational processes defined over these vehicles. We call versions of these two approaches _vehicle_ and _process_ theories of consciousness, respectively. However, while there may be space (...) for vehicle theories of consciousness in cognitive science, they are relatively rare. This is because of the influence exerted, on the one hand, by a large body of research which purports to show that the explicit representation of information in the brain and conscious experience are _dissociable_, and on the other, by the _classical_ computational theory of mind – the theory that takes human cognition to be a species of symbol manipulation. But two recent developments in cognitive science combine to suggest that a reappraisal of this situation is in order. First, a number of theorists have recently been highly critical of the experimental methodologies employed in the dissociation studies – so critical, in fact, it’s no longer reasonable to assume that the dissociability of conscious experience and explicit representation has been adequately demonstrated. Second, classicism, as a theory of human cognition, is no longer as dominant in cognitive science as it once was. It now has a lively competitor in the form of _connectionism; _and connectionism, unlike classicism, does have the computational resources to support a robust vehicle theory of consciousness. In this paper we develop and defend this connectionist vehicle theory of consciousness. It takes the form of the following simple empirical hypothesis: _phenomenal experience consists in the explicit_ _representation of information in neurally realized PDP networks_.. (shrink)

When it comes to applying computational theory to the problem of phenomenal consciousness, cognitive scientists appear to face a dilemma. The only strategy that seems to be available is one that explains consciousness in terms of special kinds of computational processes. But such theories, while they dominate the field, have counter-intuitive consequences; in particular, they force one to accept that phenomenal experience is composed of information processing effects. For cognitive scientists, therefore, it seems to come down to a choice between (...) a counter-intuitive theory or no theory at all. We offer a way out of this dilemma. We argue that the computational theory of mind doesn't force cognitive scientists to explain consciousness in terms of computational processes, as there is an alternative strategy available: one that focuses on the representational vehicles that encode information in the brain. This alternative approach to consciousness allows us to do justice to the standard intuitions about phenomenal experience, yet remain within the confines of cognitive science. (shrink)

In 1988 Fodor and Pylyshyn issued a challenge to the newly-popular connectionism: explain the systematicity of cognition without merely implementing a so-called classical architecture. Since that time quite a number of connectionist models have been put forward, either by their designers or by others, as in some measure demonstrating that the challenge can be met (e.g., Pollack, 1988, 1990; Smolensky, 1990; Chalmers, 1990; Niklasson and Sharkey, 1992; Brousse, 1993). Unfortu- nately, it has generally been unclear whether these models actually do (...) have this implication (see, for instance, the extensive philosophical debate in Smolensky, 1988; Fodor and McLaughlin, 1990; van Gelder, 1990, 1991; McLaughlin, 1993a, 1993b; Clark, 1993). Indeed, we know of no major supporter of classical orthodoxy who has felt compelled, by connectionist models and argu- ments, to concede in print that connectionists have in fact delivered a non-classical explanation of systematicity. (shrink)

Dead reckoning is a feature of the navigation behaviour shown by several creatures, including the desert ant. Recent work by C. Randy Gallistel shows that some connectionist models of dead reckoning face important challenges. These challenges are thought to arise from essential features of the connectionist approach, and have therefore been taken to show that connectionist models are unable to explain even the most primitive of psychological phenomena. I show that Gallistel’s challenges are successfully met by one recent connectionist model, (...) proposed by Ulysses Bernardet, Sergi Bermúdez i Badia, and Paul F.M.J. Verschure. The success of this model suggests that there are ways to implement dead reckoning with neural circuits that fall within the bounds of what many people regard as neurobiologically plausible, and so that the wholesale dismissal of the connectionist modelling project remains premature. (shrink)

Connectionism: Debates on Psychological Explanation, Volume Two Edited by Cynthia Macdonald and Graham Macdonald Blackwell, 1995. Pp. xvii + 424. ISBN 0-631-19744-3. 50.00 (hbk). ISBN 0-631-19745-1 16.99 (pbk). New books on the philosophy of religion Divine Hiddenness and Human Reason By J.L. Schellenberg, Cornell University Press, 1993. Pp. 217. ISBN 0-8014-2792-4. $36.50 (hbk). Reason and the Heart By William J. Wainwright, Cornell University Press, 1995. Pp. 160. ISBN 0-8014-3139-5. $28.50 (hbk). The Rationality of Belief and the Plurality of Faith Edited (...) by Thomas D. Senor, Cornell University Press, 1995. Pp. 291. ISBN 0-8014-3127-1. $39.95 (hbk). (shrink)

Explicitness has usually been approached from two points of view, labelled by Kirsh the structural and the process view, that hold opposite assumptions to determine when information is explicit. In this paper, we offer an intermediate view that retains intuitions from both of them. We establish three conditions for explicit information that preserve a structural requirement, and a notion of explicitness as a continuous dimension. A problem with the former accounts was their disconnection with psychological work on the issue. We (...) review studies by Karmiloff-Smith, and Shanks and St. John to show that the proposed conditions have psychological grounds. Finally, we examine the problem of explicit rules in connectionist systems in the light of our framework. (shrink)

Explicitness has usually been approached from two points of view, labelled by Kirsh the structural and the process view, that hold opposite assumptions to determine when information is explicit. In this paper, we offer an intermediate view that retains intuitions from both of them. We establish three conditions for explicit information that preserve a structural requirement, and a notion of explicitness as a continuous dimension. A problem with the former accounts was their disconnection with psychological work on the issue. We (...) review studies by Karmiloff-Smith, and Shanks and St. John to show that the proposed conditions have psychological grounds. Finally, we examine the problem of explicit rules in connectionist systems in the light of our framework. (shrink)

One of the best known arguments against the connectionist approach to artificial intelligence and cognitive science is that neural networks are black boxes, i.e., there is no understandable account of their operation. This difficulty has impeded efforts to explain how categories arise from raw sensory data. Moreover, it has complicated investigation about the role of symbols and language in cognition. This state of things has been radically changed by recent experimental findings in artificial deep learning research. Two kinds of artificial (...) deep learning networks, namely the convolutional neural network and the generative adversarial network have been found to possess the capability to build internal states that are interpreted by humans as complex visual categories, without any specific hints or any grammatical processing. This emergent ability suggests that those categories do not depend on human knowledge or the syntactic structure of language, while they do rely on their visual context. This supports a mild form of empiricism, while it does not assume that computational functionalism is true. Some consequences are extracted regarding the debate about amodal and grounded representations in the human brain. Furthermore, new avenues for research on cognitive science are open. (shrink)

Connectionism and phenomenology can mutually inform and mutually constrain each other. In this manifesto I outline an approach to consciousness based on distinctions developed by connectionists. Two core identities are central to a connectionist theory of consciouness: conscious states of mind are identical to occurrent activation patterns of processing units; and the variable dispositional strengths on connections between units store latent and unconscious information. Within this broad framework, a connectionist model of consciousness succeeds according to the degree of correspondence between (...) the content of human consciousness (the world as it is experienced) and the interpreted content of the network. Constitutive self-awareness and reflective self-awareness can be captured in a model through its ability to respond to self-reflexive information, identify self-referential categories, and process information in the absence of simultaneous input. The qualitative feel of sensation appears in a model as states of activation that are not fully discriminated by later processing. Connectionism also uniquely explains several specific features of experience. The most important of these is the superposition of information in consciousness — our ability to perceive more than meets the eye, and to apprehend complex categorical and temporal information in a single highly-cognized glance. This superposition in experience matches a superposition of representational content in distributed representations. (shrink)

Accounting for phenomenal structure—the forms, aspects, and features of conscious experience—poses a deep challenge for the scientific study of consciousness, but rather than abandon hope I propose a way forward. Connectionism, I argue, offers a bi-directional analogy, with its oft-noted “neural inspiration” on the one hand, and its largely unnoticed capacity to illuminate our phenomenology on the other. Specifically, distributed representations in a recurrent network enable networks to superpose categorical, contextual, and temporal information on a specific input representation, much as (...) our own experience does. Artificial neural networks also suggest analogues of four salient distinctions between sensory and nonsensoty consciousness. The paper concludes with speculative proposals for discharging the connectionist heuristics to leave a robust, detailed empirical theory of consciousness. (shrink)

Abstract In this paper we address three concepts that are much talked about in the animal robotics community. These concepts are (1) representations, (2) goals, and (3) minimal cognition. We want to distinguish between information as an objective commodity and representation as something which involves a user, i.e. a system which accesses and uses information. Information per se lies out there and exists independently of any system that makes use of it. Representations presuppose design and require a user. We want (...) to distinguish different kinds of purposive behaviour, which involve different notions of goals, which themselves appear to be different from the point of view of the agent, the designer and the naive observer. We want to try to define cognition in terms of the minimal criteria that would be acceptable in view of our definitions of representations and goals. We recognise that full?blown cognition may require ingredients other than those of minimal cognition. (shrink)

As AI systems become increasingly complex it may become unclear, even to the designer of a system, why exactly a system does what it does. This leads to a lack of trust in AI systems. To solve this, the field of explainable AI has been working on ways to produce explanations of these systems’ behaviors. Many methods in explainable AI, such as LIME, offer only a statistical argument for the validity of their explanations. However, some methods instead study the internal (...) structure of the system and try to find components which can be assigned an interpretation. I believe that these methods provide more valuable explanations than those statistical in nature. I will try to identify which explanations can be considered internal to the system using the Chomskyan notion of tacit knowledge. I argue that each explanation expresses a rule, and through the localization of this rule in the system internals, we can take a system to have tacit knowledge of the rule. I conclude that the only methods which are able to sufficiently establish this tacit knowledge are those along the lines of Olah : 4901–4911, 2017), and therefore they provide explanations with unique strengths. (shrink)

Stephen Laurence and Eric Margolis have recently argued that certain kinds of regress arguments against the language of thought (LOT) hypothesis as an account of how we understand natural languages have been answered incorrectly or inadequately by supporters of LOT ('Regress arguments against the language of thought', Analysis, 57 (1), 60-6, J 97). They argue further that this does not undermine the LOT hypothesis, since the main sources of support for LOT are (or might be) independent of it providing an (...) account of how we understand natural language. In my paper I seek to refute both these claims, and reinstate the putative explanation of natural language understanding as a necessarily central part of the support for LOT. The main argument exploits the fact that Laurence and Margolis give too little weight to the ideas (a) that LOT might be innate (b) that for LOT supporters a semantic theory must apply to in-the-head tokens, not linguistic utterances. (shrink)

In this paper, I seek to refute arguments for the idea that folk psychological explanation, i.e., the explanation of actions, beliefs and desires in terms of one another, should be understood as being of a different character than ordinary scientific explanations, a view defended most prominently in analytical philosophy by Donald Davidson and John McDowell. My strategy involves arguing both against the extant arguments for the idea that FP must be construed as giving such explanations, and also against the very (...) notion of such a different kind of explanation. I argue first that the in some sense a priori and conceptual nature of folk psychological principles does not support the idea that these are other than empirical generalisations, by appeal to recent nativist ideas in cognitive science and to Lewis's conception of the meaning of theoretical terms. Second, I argue that there is no coherent sense in which folk psychological explanations can be seen as normative. Thirdly, I examine the putatively holistic character of the mental and conclude that that too fails to provide any cogent reasons for viewing folk psychological explanations as different from other kinds of explanation. (shrink)

I argue that intentional psychology does not stand in need of vindication by a lower-level implementation theory from cognitive science, in particular the representational theory of mind (RTM), as most famously Jerry Fodor has argued. The stance of the paper is novel in that I claim this holds even if one, in line with Fodor, views intentional psychology as an empirical theory, and its theoretical posits as as real as those of other sciences. I consider four metaphysical arguments for the (...) idea that intentional psychological states, such as beliefs, must be seen as requiring in-the-head mental representations for us to be able to understand their characteristic causal powers and argue that none of them validly generate their desired conclusions. I go on to argue that RTM, or some computational version thereof, is not motivated by appeal to the nature of cognitive science research either. I conclude that intentional psychology, though an empirical theory, is autonomous from details of lower level mechanism in a way that renders RTM unwarranted. (shrink)