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)

Deep learning is currently the most prominent and widely successful method in artificial intelligence. Despite having played an active role in earlier artificial intelligence and neural network research, philosophers have been largely silent on this technology so far. This is remarkable, given that deep learning neural networks have blown past predicted upper limits on artificial intelligence performance—recognizing complex objects in natural photographs and defeating world champions in strategy games as complex as Go and chess—yet there remains no universally accepted explanation (...) as to why they work so well. This article provides an introduction to these networks as well as an opinionated guidebook on the philosophical significance of their structure and achievements. It argues that deep learning neural networks differ importantly in their structure and mathematical properties from the shallower neural networks that were the subject of so much philosophical reflection in the 1980s and 1990s. The article then explores several different explanations for their success and ends by proposing three areas of inquiry that would benefit from future engagement by philosophers of mind and science. (shrink)

Deep neural networks have become increasingly successful in applications from biology to cosmology to social science. Trained DNNs, moreover, correspond to models that ideally allow the prediction of new phenomena. Building in part on the literature on ‘eXplainable AI’, I here argue that these models are instrumental in a sense that makes them non-explanatory, and that their automated generation is opaque in a unique way. This combination implies the possibility of an unprecedented gap between discovery and explanation: When unsupervised models (...) are successfully used in exploratory contexts, scientists face a whole new challenge in forming the concepts required for understanding underlying mechanisms. (shrink)

The ongoing explosion of interest in artificial intelligence is fueled in part by recently developed techniques in machine learning. Those techniques allow automated systems to process huge amounts of data, utilizing mathematical methods that depart from traditional statistical approaches, and resulting in impressive advancements in our ability to make predictions and uncover correlations across a host of interesting domains. But as is now widely discussed, the way that those systems arrive at their outputs is often opaque, even to the experts (...) who design and deploy them. Is it morally problematic to make use of opaque automated methods when making high-stakes decisions, like whether to issue a loan to an applicant, or whether to approve a parole request? Many scholars answer in the affirmative. However, there is no widely accepted explanation for why transparent systems are morally preferable to opaque systems. We argue that the use of automated decision-making systems sometimes violates duties of consideration that are owed by decision-makers to decision-subjects, duties that are both epistemic and practical in character. Violations of that kind generate a weighty consideration against the use of opaque decision systems. In the course of defending our approach, we show that it is able to address three major challenges sometimes leveled against attempts to defend the moral import of transparency in automated decision-making. (shrink)

Some machine learning models, in particular deep neural networks (DNNs), are not very well understood; nevertheless, they are frequently used in science. Does this lack of understanding pose a problem for using DNNs to understand empirical phenomena? Emily Sullivan has recently argued that understanding with DNNs is not limited by our lack of understanding of DNNs themselves. In the present paper, we will argue, _contra_ Sullivan, that our current lack of understanding of DNNs does limit our ability to understand with (...) DNNs. Sullivan’s claim hinges on which notion of understanding is at play. If we employ a weak notion of understanding, then her claim is tenable, but rather weak. If, however, we employ a strong notion of understanding, particularly explanatory understanding, then her claim is not tenable. (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)

Can humans and artificial intelligences share concepts and communicate? Making AI Intelligible shows that philosophical work on the metaphysics of meaning can help answer these questions. Herman Cappelen and Josh Dever use the externalist tradition in philosophy to create models of how AIs and humans can understand each other. In doing so, they illustrate ways in which that philosophical tradition can be improved. The questions addressed in the book are not only theoretically interesting, but the answers have pressing practical implications. (...) Many important decisions about human life are now influenced by AI. In giving that power to AI, we presuppose that AIs can track features of the world that we care about (for example, creditworthiness, recidivism, cancer, and combatants). If AIs can share our concepts, that will go some way towards justifying this reliance on AI. This ground-breaking study offers insight into how to take some first steps towards achieving Interpretable AI. (shrink)

The last 5 years have seen a series of remarkable achievements in deep-neural-network-based artificial intelligence research, and some modellers have argued that their performance compares favourably to human cognition. Critics, however, have argued that processing in deep neural networks is unlike human cognition for four reasons: they are (i) data-hungry, (ii) brittle, and (iii) inscrutable black boxes that merely (iv) reward-hack rather than learn real solutions to problems. This article rebuts these criticisms by exposing comparative bias within them, in the (...) process extracting some more general lessons that may also be useful for future debates. (shrink)

In March 2016, DeepMind's computer programme AlphaGo surprised the world by defeating the world‐champion Go player, Lee Sedol. AlphaGo exhibits a novel, surprising and valuable style of play and has been recognised as “creative” by the artificial intelligence (AI) and Go communities. This article examines whether AlphaGo engages in creative problem solving according to the standards of comparative psychology. I argue that AlphaGo displays one important aspect of creative problem solving (namely mental scenario building in the form of Monte Carlo (...) tree search), while lacking another (domain generality). This analysis has consequences for how we think about creativity in humans and AI. (shrink)

I will discuss a variety of cases such that the subject's believing truly is somewhat of an accident, but less so than in a Gettier case. In each case, this is because her reasons are not ultimately undefeated full stop, but they are ultimately undefeated with certain qualifications. For example, the subject's reasons might be ultimately defeated considered in themselves but ultimately undefeated considered as a proper part of an inference to the best explanation that is undefeated without qualification. In (...) each section of the paper, I consider different qualifications and show how in each case we get an epistemic standing I call “coach-class knowledge”. First-class knowledge requires justifying reasons that are undefeated without qualification. Coach-class knowledge only requires qualified lack of defeat. I will use this distinction to bring debates over knowledge from falsehood and fake barns to an ecumenical resolution. In both cases, the subject enjoys coach-class rather than first-class knowledge. I will also show that the defeasible reasoning tradition can better account for graded accidental truth than safety theories. (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)

There is a vast literature within philosophy of mind that focuses on artificial intelligence, but hardly mentions methodological questions. There is also a growing body of work in philosophy of science about modeling methodology that hardly mentions examples from cognitive science. Here these discussions are connected. Insights developed in the philosophy of science literature about the importance of idealization provide a way of understanding the neural implausibility of connectionist networks. Insights from neurocognitive science illuminate how relevant similarities between models and (...) targets are picked out, how modeling inferences are justified, and the metaphysical status of models. (shrink)

This paper investigates how intuitions about scientific discovery using artificial intelligence can be used to improve our understanding of scientific discovery more generally. Traditional accounts of discovery have been agent-centred: they place emphasis on identifying a specific agent who is responsible for conducting all, or at least the important part, of a discovery process. We argue that these accounts experience difficulties capturing scientific discovery involving AI and that similar issues arise for human discovery. We propose an alternative, collective-centred view as (...) superior for understanding discovery, with and without AI. This view maintains that discovery is performed by a collective of agents and entities, each making contributions that differ in significance and character, and that attributing credit for discovery depends on various finer-grained properties of the contributions made. Detailing its conceptual resources, we argue that this view is considerably more compelling than its agent-centred alternative. Considering and responding to several theoretical and practical challenges, we point to concrete avenues for further developing the view we propose. (shrink)

The concept of “representation” is used broadly and uncontroversially throughout neuroscience, in contrast to its highly controversial status within the philosophy of mind and cognitive science. In this paper I first discuss the way that the term is used within neuroscience, in particular describing the strategies by which representations are characterized empirically. I then relate the concept of representation within neuroscience to one that has developed within the field of machine learning. I argue that the recent success of artificial neural (...) networks on certain tasks such as visual object recognition reflects the degree to which those systems exhibit inherent inductive biases that reflect the structure of the physical world. I further argue that any system that is going to behave intelligently in the world must contain representations that reflect the structure of the world; otherwise, the system must perform unconstrained function approximation which is destined to fail due to the curse of dimensionality, in which the number of possible states of the world grows exponentially with the number of dimensions in the space of possible inputs. An analysis of these concepts in light of philosophical debates regarding the ontological status of representations suggests that the representations identified within both biological and artificial neural networks qualify as legitimate representations in the philosophical sense. (shrink)

Some embodied theories of concepts state that concepts are represented in a sensorimotor manner, typically via simulation in sensorimotor cortices. Fred Adams (2010) has advanced an empirical argument against embodied concepts reasoning as follows. If concepts are embodied, then patients with certain sensorimotor impairments should perform worse on categorization tasks involving those concepts. Adams cites a study with Moebius Syndrome patients that shows typical categorization performance in face-based emotion recognition. Adams concludes that their typical performance shows that embodiment is false. (...) Moebius patients must draw on amodal (non-embodied) emotion concepts. In this paper, I review face-based emotion recognition studies with Moebius patients yielding conflicting results and diagnose these conflicts as a difference in experimental design. When emotion labels are provided, patients have typical performance, but when labels are not provided patients are severely deficient. I then show how an embodied, psychological constructionist view of emotions predicts and explains these performance differences. The upshot is that embodied theories of concepts are vindicated. (shrink)

The use of machine learning instead of traditional models in neuroscience raises significant questions about the epistemic benefits of the newer methods. I draw on the literature on model intelligibility in the philosophy of science to offer some benchmarks for the interpretability of artificial neural networks used as a predictive tool in neuroscience. Following two case studies on the use of ANN’s to model motor cortex and the visual system, I argue that the benefit of providing the scientist with understanding (...) of the brain trades off against the predictive accuracy of the models. This trade-off between prediction and understanding is better explained by a non-factivist account of scientific understanding. (shrink)

This paper traces the empiricist program from early debates between nativism and behaviorism within philosophy, through debates about early connectionist approaches within the cognitive sciences, and up to their recent iterations within the domain of deep learning. We demonstrate how current debates on the nature of cognition via deep network architecture echo some of the core issues from the Chomsky/Quine debate and investigate the strength of support offered by these various lines of research to the empiricist standpoint. Referencing literature from (...) both computer science and philosophy, we conclude that the current state of deep learning does not offer strong encouragement to the empiricist side despite some arguments to the contrary. (shrink)

According to a recent proposal, natural concepts are represented in an optimally designed similarity space, adhering to principles a skilled engineer would use for creatures with our perceptual and cognitive capacities. One key principle is that natural concepts should be easily learnable. While evidence exists for parts of this optimal design proposal, there has been no direct evidence linking naturalness to learning until now. This article presents results from a computational study on perceptual color space, demonstrating that naturalness indeed facilitates (...) learning. (shrink)

A central task of developmental psychology and philosophy of science is to show how humans learn radically new concepts. Famously, Fodor has argued that such learning is impossible if concepts have definitional structure and all learning is hypothesis testing. We present several learning processes that can generate novel concepts. They yield transformations of the fundamental feature space, generating new similarity structures which can underlie conceptual change. This framework provides a tractable, empiricist-friendly account that unifies and shores up various strands of (...) the neo-Quinean approach to conceptual development. (shrink)

A basic deep neural network (DNN) is trained to exhibit a large set of input–output dispositions. While being a good model of the way humans perform some tasks automatically, without deliberative reasoning, more is needed to approach human‐like artificial intelligence. Analysing recent additions brings to light a distinction between two fundamentally different styles of computation: content‐specific and non‐content‐specific computation (as first defined here). For example, deep episodic RL networks draw on both. So does human conceptual reasoning. Combining the two takes (...) advantage of the complementary costs and benefits of each. It also offers a better model of human cognitive competence. (shrink)

A novel account of semantic information is proposed. The gist is that structural correspondence, analysed in terms of similarity, underlies an important kind of semantic information. In contrast to extant accounts of semantic information, it does not rely on correlation, covariation, causation, natural laws, or logical inference. Instead, it relies on structural similarity, defined in terms of correspondence between classifications of tokens into types. This account elucidates many existing uses of the notion of information, for example, in the context of (...) scientific models and structural representations in cognitive science. It is poised to open a new research program concerned with various kinds of semantic information, its functions, and its measurement. (shrink)

Recent scholarship emphasizes the scaffolding role of language for cognition. Language, it is claimed, is a cognition-enhancing niche ( Clark, 2006 ), a programming tool for cognition ( Lupyan and Bergen, 2016 ), even neuroenhancement ( Dove, 2019 ) and augments cognitive functions such as memory, categorization, cognitive control, and meta-cognitive abilities (“thinking about thinking”). Yet, the notion that language enhances or augments cognition, and in particular, cognitive control does not easily fit in with embodied approaches to language processing, or (...) so we will argue. Accounts aiming to explain how language enhances various cognitive functions often employ a notion of abstract representation. Yet, embodied approaches to language processing have it that language processing crucially, according to some accounts even exclusively, involves embodied, modality-specific, i.e., non-abstract representations. In coming to understand a particular phrase or sentence, a prior experience has to be simulated or reenacted. The representation thus activated is embodied (modality-specific) as sensorimotor regions of the brain are thereby recruited. In this paper, we will first discuss the notion of representation, clarify what it takes for a representation to be embodied or abstract, and distinguish between conceptual and (other) linguistic representations. We will then put forward a characterization of cognitive control and examine its representational infrastructure. The remainder of the paper will be devoted to arguing that language augments cognitive control. To that end, we will draw on two lines of research, which investigate how language augments cognitive control: (i) research on the availability of linguistic labels and (ii) research on the active usage of a linguistic code, specifically, in inner speech. Eventually, we will argue that the cognition-enhancing capacity of language can be explained once we assume that it provides us with (a) abstract, non-embodied representations and with (b) abstract, sparse linguistic representations that may serve as easy-to-manipulate placeholders for fully embodied or otherwise more detailed representations. (shrink)

The goal of the paper is to develop and propose a general model of the state space of AI. Given the breathtaking progress in AI research and technologies in recent years, such conceptual work is of substantial theoretical interest. The present AI hype is mainly driven by the triumph of deep learning neural networks. As the distinguishing feature of such networks is the ability to self-learn, self-learning is identified as one important dimension of the AI state space. Another dimension is (...) recognized as generalization, the possibility to go over from specific to more general types of problems. A third dimension is semantic grounding. Our overall analysis connects to a number of known foundational issues in the philosophy of mind and cognition: the blockhead objection, the Turing test, the symbol grounding problem, the Chinese room argument, and use theories of meaning. It shall finally be argued that the dimension of grounding decomposes into three sub-dimensions. And the dimension of self-learning turns out as only one of a whole range of “self-x-capacities” (based on ideas of organic computing) that span the self-x-subspace of the full AI state space. (shrink)

This article argues that the algorithms known as neural nets underlie a new form of artificial intelligence that we call indexical AI. Contrasting with the once dominant symbolic AI, large-scale learning systems have become a semiotic infrastructure underlying global capitalism. Their achievements are based on a digital version of the sign-function index, which points rather than describes. As these algorithms spread to parse the increasingly heavy data volumes on platforms, it becomes harder to remain skeptical of their results. We call (...) social faith in these systems the naive iconic interpretation of AI and position their indexical function between heuristic symbol use and real intelligence, opening the black box to reveal semiotic function. (shrink)

To learn about real world phenomena, scientists have traditionally used models with clearly interpretable elements. However, modern machine learning (ML) models, while powerful predictors, lack this direct elementwise interpretability (e.g. neural network weights). Interpretable machine learning (IML) offers a solution by analyzing models holistically to derive interpretations. Yet, current IML research is focused on auditing ML models rather than leveraging them for scientific inference. Our work bridges this gap, presenting a framework for designing IML methods—termed ’property descriptors’—that illuminate not just (...) the model, but also the phenomenon it represents. We demonstrate that property descriptors, grounded in statistical learning theory, can effectively reveal relevant properties of the joint probability distribution of the observational data. We identify existing IML methods suited for scientific inference and provide a guide for developing new descriptors with quantified epistemic uncertainty. Our framework empowers scientists to harness ML models for inference, and provides directions for future IML research to support scientific understanding. (shrink)

Artificial intelligent (AI) systems that perform image classification tasks are being used to great success in many application contexts. However, many of these systems are opaque, even to experts. This lack of understanding can be problematic for ethical, legal, or practical reasons. The research field Explainable AI (XAI) has therefore developed several approaches to explain image classifiers. The hope is to bring about understanding, e.g., regarding why certain images are classified as belonging to a particular target class. Most of these (...) approaches use visual explanations. Drawing on Elgin’s work (True enough. MIT Press, Cambridge, 2017), I argue that analyzing what those explanations exemplify can help to assess their suitability for producing understanding. More specifically, I suggest to distinguish between two forms of examples according to their suitability for producing understanding. I call these forms samples and exemplars, respectively. Samples are prone to misinterpretation and thus carry the risk of leading to misunderstanding. Exemplars, by contrast, are intentionally designed or chosen to meet contextual requirements and to mitigate the risk of misinterpretation. They are thus preferable for bringing about understanding. By reviewing several XAI approaches directed at image classifiers, I show that most of them explain with samples. If my analysis is correct, it will be beneficial if such explainability methods use explanations that qualify as exemplars. (shrink)

: Fodor’s view of the mind is thoroughly computational. This means that the basic kind of mental entity is a “discursive” mental representation and operations over this kind of mental representation have broad architectural scope, extending out to the edges of perception and the motor system. However, in multiple epochs of his work, Fodor attempted to define a functional role for non-discursive, imagistic representation. I describe and critique his two considered proposals. The first view says that images play a particular (...) kind of functional role in certain types of deliberative tasks. The second says that images are solely restricted to the borders of perception, and act as a sort of medium for the fixing of conceptual reference. I argue, against the first proposal, that a broad-scope computationalism such as Fodor’s renders images in principle functionally redundant. I argue, against the second proposal, that empirical evidence suggests that non-discursive representations are learned through perceptual learning, and directly inform category judgments. In each case, I point out extant debates for which the arguments are relevant. The upshot is that there is motivation for limited scope computationalism, in which some, but not all, mental processes operate on discursive mental representations. Keywords: Computational Theory of Mind; Mental Representation; Perception; Mental Image; Jerry Fodor Fodor e le rappresentazioni mentali come immagini Riassunto: La concezione della mente di Fodor è rigorosamente computazionale, ossia le entità mentali di base sono rappresentazioni mentali “discorsive”. Le operazioni su queste rappresentazioni hanno un fine architettonico ampio, che va fino ai confini della percezione e del sistema motorio. In periodi diversi del suo lavoro, Fodor ha proposto due modi per definire un ruolo funzionale per la rappresentazione non-discorsiva come immagine. Tratterò criticamente entrambi. Per il primo, le immagini giocano un particolare tipo di ruolo funzionale in certi tipi di compiti deliberativi, mentre, per il secondo, sono relegate unicamente ai confini della percezione, agendo come medium per fissare il riferimento concettuale. Contro il primo sosterrò che un computazionalismo così ampio come quello di Fodor rende le immagini in principio funzionalmente ridondanti. Contro il secondo sosterrò che l’evidenza empirica suggerisce che le rappresentazioni non-discorsive vengono apprese percettivamente, agendo direttamente sui giudizi di categorizzazione. In entrambi i casi considererò gli argomenti più rilevanti nel dibattito corrente. Si vedrà che ci sono buone ragioni in favore di un computazionalismo più limitato, in cui alcuni processi mentali operano su rappresentazioni mentali discorsive. Parole chiave: Teoria computazionale della mente; Rappresentazione mentale; Percezione; Immagine mentale; Jerry Fodor. (shrink)

In this article, M. Beatrice Fazi takes up Media Theory’s invitation to engage with Alan Díaz Alva’s analysis of her philosophical work on contingency in computation. The central argument of Fazi’s Contingent Computation: Abstraction, Experience, and Indeterminacy in Computational Aesthetics is that computation can be productive of ontological novelty. This piece revisits that argument in the light of the technological developments that have occurred since 2018, when the book was published. Focusing on generative artificial intelligence (generative AI), the article considers (...) how the concepts of generativity, aesthetics and contingency are tackled in the book and in present-day debates about generative AI systems. (shrink)

Connectionist research first emerged in the 1940s. The first phase of connectionism attracted a certain amount of media attention, but scant philosophical interest. The phase came to an abrupt halt, due to the efforts of Minsky and Papert (1969), when they argued for the intrinsic limitations of the approach. In the mid-1980s connectionism saw a resurgence. This marked the beginning of the second phase of connectionist research. This phase did attract considerable philosophical attention. It was of philosophical interest, as it (...) offered a way of counteracting the conceptual ties to the philosophical traditions of atomism, rationalism, logic, nativism, rule realism and a concern with the role symbols play in human cognitive functioning, which was prevalent as a consequence of artificial intelligence research. The surge in philosophical interest waned, possibly in part due to the efforts of some traditionalists and the so-called black box problem. Most recently, what may be thought of as a third phase of connectionist research, based on so-called deep learning methods, is beginning to show some signs of again exciting philosophical interest. (shrink)

This paper analyzes responsible human–machine interaction concerning artificial neural networks (ANNs) and large language models (LLMs) by considering the extension of human agency and autonomy by means of artificial intelligence (AI). Thereby, the paper draws on the sociological concept of “interobjectivity,” first introduced by Bruno Latour, and applies it to technologically situated and interconnected agency. Drawing on Don Ihde’s phenomenology of human-technology relations, this interobjective account of AI allows to understand human–machine interaction as embedded in the social lifeworld. Finally, the (...) paper develops a connectionist account of responsible AI, thereby focusing on patterns such as social goals and actions. (shrink)

It has often been argued that we face a trade-off between accuracy and opacity in deep learning models. The idea is that we can only harness the accuracy of deep learning models by simultaneously accepting that the grounds for the models’ decision-making are epistemically opaque to us. In this paper, we ask the following question: what are the prospects of making deep learning models transparent without compromising on their accuracy? We argue that the answer to this question depends on which (...) kind of opacity we have in mind. If we focus on the standard notion of opacity, which tracks the internal complexities of deep learning models, we argue that existing explainable AI (XAI) techniques show us that the prospects look relatively good. But, as it has recently been argued in the literature, there is another notion of opacity that concerns factors external to the model. We argue that there are at least two types of external opacity—link opacity and structure opacity—and that existing XAI techniques can to some extent help us reduce the former but not the latter. (shrink)

The aim of this paper is to grasp the relevant distinctions between various ways in which models and simulations in Artificial Intelligence (AI) relate to cognitive phenomena. In order to get a systematic picture, a taxonomy is developed that is based on the coordinates of formal versus material analogies and theory-guided versus pre-theoretic models in science. These distinctions have parallels in the computational versus mimetic aspects and in analytic versus exploratory types of computer simulation. The proposed taxonomy cuts across the (...) traditional dichotomies between symbolic and embodied AI, general intelligence and symbol and intelligence and cognitive simulation and human/non-human-like AI. -/- According to the taxonomy proposed here, one can distinguish between four distinct general approaches that figured prominently in early and classical AI, and that have partly developed into distinct research programs: first, phenomenal simulations (e.g., Turing’s “imitation game”); second, simulations that explore general-level formal isomorphisms in pursuit of a general theory of intelligence (e.g., logic-based AI); third, simulations as exploratory material models that serve to develop theoretical accounts of cognitive processes (e.g., Marr’s stages of visual processing and classical connectionism); and fourth, simulations as strictly formal models of a theory of computation that postulates cognitive processes to be isomorphic with computational processes (strong symbolic AI). -/- In continuation of pragmatic views of the modes of modeling and simulating world affairs, this taxonomy of approaches to modeling in AI helps to elucidate how available computational concepts and simulational resources contribute to the modes of representation and theory development in AI research—and what made that research program uniquely dependent on them. (shrink)

The ongoing explosion of interest in artificial intelligence is fueled in part by recently developed techniques in machine learning. Those techniques allow automated systems to process huge amounts of data, utilizing mathematical methods that depart from traditional statistical approaches, and resulting in impressive advancements in our ability to make predictions and uncover correlations across a host of interesting domains. But as is now widely discussed, the way that those systems arrive at their outputs is often opaque, even to the experts (...) who design and deploy them. Is it morally problematic to make use of opaque automated methods when making high-stakes decisions, like whether to issue a loan to an applicant, or whether to approve a parole request? Many scholars answer in the affirmative. However, there is no widely accepted explanation for why transparent systems are morally preferable to opaque systems. We argue that the use of automated decision-making systems sometimes violates duties of consideration that are owed by decision-makers to decision-subjects, duties that are both epistemic and practical in character. Violations of that kind generate a weighty consideration against the use of opaque decision systems. In the course of defending our approach, we show that it is able to address three major challenges sometimes leveled against attempts to defend the moral import of transparency in automated decision-making. (shrink)

In this essay, it is argued that naturalism of an even moderate sort speaks strongly against a certain widely held thesis about the human mental (and cognitive) architecture: that it is divided into two distinct levels, the personal and the subpersonal, about the former of which we gain knowledge in a manner that effectively insulates such knowledge from the results of scientific research. -/- An empirically motivated alternative is proposed, according to which the architecture is, so to speak, flattened from (...) above. On this flattened view, although the states and processes typically associated with the personal level likely appear in our best models of the production of human behavior, they appear alongside states and processes normally associated with the subpersonal level. Moreover, the success of such models depends nowise on a levels-based distinction between the various causal contributors. It is argued that the flattened view has methodological implications of significant import. (shrink)

Despite the recent upsurge in research on abstract concepts, there remain puzzles at the foundation of their empirical study. These are most evident when we consider what is required to assess a person’s abstract conceptual abilities without using language as a prompt or requiring it as a response—as in classic non-verbal categorization tasks, which are standardly considered tests of conceptual understanding. After distinguishing two divergent strands in the most common conception of what it is for a concept to be abstract, (...) we argue that neither reliably captures the kind of abstraction required to successfully categorize in non-verbal tasks. We then present a new conception of concept abstractness—termed Trial Concreteness—that is keyed to individual categorization trials. It has advantages in capturing the context-relativity of the degree of abstraction required for the application of a concept and fittingly correlates with participant success in recent experiments. (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)

Common everyday materials such as textiles, foodstuffs, soil or skin can have complex, mutable and varied appearances. Under typical viewing conditions, most observers can visually recognize materials effortlessly, and determine many of their properties without touching them. Visual material perception raises many fascinating questions for vision researchers, neuroscientists and philosophers, yet has received little attention compared to the perception of color or shape. Here we discuss some of the challenges that material perception raises and argue that further philosophical thought should (...) be directed to how we see materials. (shrink)

The representations of deep convolutional neural networks (CNNs) are formed from generalizing similarities and abstracting from differences in the manner of the empiricist theory of abstraction (Buckner, Synthese 195:5339–5372, 2018). The empiricist theory of abstraction is well understood to entail infinite regress and circularity in content constitution (Husserl, Logical Investigations. Routledge, 2001). This paper argues these entailments hold a fortiori for deep CNNs. Two theses result: deep CNNs require supplementation by Quine’s “apparatus of identity and quantification” in order to (1) (...) achieve concepts, and (2) represent objects, as opposed to “half-entities” corresponding to similarity amalgams (Quine, Quintessence, Cambridge, 2004, p. 107). Similarity amalgams are also called “approximate meaning[s]” (Marcus & Davis, Rebooting AI, Pantheon, 2019, p. 132). Although Husserl inferred the “complete abandonment of the empiricist theory of abstraction” (a fortiori deep CNNs) due to the infinite regress and circularity arguments examined in this paper, I argue that the statistical learning of deep CNNs may be incorporated into a Fodorian hybrid account that supports Quine’s “sortal predicates, negation, plurals, identity, pronouns, and quantifiers” which are representationally necessary to overcome the regress/circularity in content constitution and achieve objective (as opposed to similarity-subjective) representation (Burge, Origins of Objectivity. Oxford, 2010, p. 238). I base myself initially on Yoshimi’s (New Frontiers in Psychology, 2011) attempt to explain Husserlian phenomenology with neural networks but depart from him due to the arguments and consequently propose a two-system view which converges with Weiskopf’s proposal (“Observational Concepts.” The Conceptual Mind. MIT, 2015. 223–248). (shrink)

A typical expository strategy for the predictive processing account begins with perception and then extends to other cognitive domains, such as action or non-human animal cognition. Because this standard, perception-first expository strategy begins at the end of an evolutionary process, it may introduce both diachronic and synchronic distortions into the overall account. As far as the diachronic distortion is concerned, because the perception-first strategy presupposes a highly decoupled cognitive architecture, it invites us to project this architecture onto the coupled cognitive (...) systems of our evolutionary ancestors and other non-human organisms. Turning to the synchronic distortion, as the perception-first strategy is extended to conative states, it invites a cognitivism that seeks to reconstrue motivational states as a species of belief. In this paper, I explore what a predictive processing account would look like if it began at the beginning, with the highly coupled representations carried by our evolutionary ancestors and other non-human organisms. I conclude by exploring the proposal’s implications for a non-cognitivist, predictive processing account of action. (shrink)

Artificial neural networks and supervised learning have become an essential part of science. Beyond using them for accurate input-output mapping, there is growing attention to a new feature-oriented approach. Under the assumption that networks optimised for a task may have learned to represent and utilise important features of the target system for that task, scientists examine how those networks manipulate inputs and employ the features networks capture for scientific discovery. We analyse this approach, show its hidden caveats, and suggest its (...) legitimate use. We distinguish three things that scientists call a ‘feature’: parametric, diagnostic, and real-world features. The feature-oriented approach aims for real-world features by interpreting the former two, which also partially rely on the network. We argue that this approach faces a problem of non-uniqueness: there are numerous discordant parametric and diagnostic features and ways to interpret them. When the approach aims at novel discovery, scientists often need to choose between those options, but they lack the background knowledge to justify their choices. Consequentially, features thus identified are not promised to be real. We argue that they should not be used as evidence but only used instrumentally. We also suggest transparency in feature selection and the plurality of choices. (shrink)

Yoshimi has attempted to defuse my argument concerning the identification of network abstraction with empiricist abstraction - thus entailing psychologism - by claiming that the argument does not generalize from the example of simple feed-forward networks. I show that the particular details of networks are logically irrelevant to the nature of the abstractive process they employ. This is ultimately because deep artificial neural networks (ANNs) and dynamical systems theory applied to the mind (DST) are both associationisms - that is, empiricist (...) theories that derive the principles of thought from the causal history of the organism/system. On this basis, I put forward a new aspect of the old argument by noting that ANNs & DST are the causal bases of the phenomena of passive synthesis, whereas the language of thought hypothesis (LOT) and the symbolic computational theory of mind (CTM) are the causal bases of the phenomena of active synthesis. If the phenomena of active synthesis are not distinct in kind from and are thus reducible to those of passive synthesis, psychologism results. Yoshimi’s program, insofar as it denies this fundamental phenomenological distinction, is revealed to be the true anti-pluralist program, by essentially denying the causal efficacy of the mechanistic foundations of active synthesis by referring phenomenology exclusively to associationism for its causal foundation. (shrink)

This paper aims to offer a new view of the role of connectionist models in the study of human cognition through the conceptualization of the history of connectionism – from the simplest perceptrons to convolutional neural nets based on deep learning techniques, as well as through the interpretation of criticism coming from symbolic cognitive science. Namely, the connectionist approach in cognitive science was the target of sharp criticism from the symbolists, which on several occasions caused its marginalization and almost complete (...) abandonment of its assumptions in the study of cognition. Criticisms have mostly pointed to its explanatory inadequacy as a theory of cognition or to its biological implausibility as a theory of implementation, and critics often focused on specific shortcomings of some connectionist models and argued that they apply to connectionism in general. In this paper, we want to show that both types of critique are based on the assumption that the only valid explanations in cognitive science are instances of homuncular functionalism and that by removing this assumption and by adopting an alternative methodology – exploratory mechanistic strategy, we can reject most objections to connectionism as irrelevant, explain the progress of connectionist models despite their shortcomings and sketch the trajectory of their future development. By adopting mechanistic explanations and by criticizing functionalism, we will reject the objections of explanatory inadequacy, by characterizing connectionist models as generic rather than concrete mechanisms, we will reject the objections of biological implausibility, and by attributing the exploratory character to connectionist models we will show that practice of generalizing current to general failures of connectionism is unjustified. (shrink)

Cognition, it is often heard nowadays, is embodied. My concern is with embodied accounts of language comprehension. First, the basic idea will be outlined and some of the evidence that has been put forward in their favor will be examined. Second, their empiricist heritage and their conception of abstract ideas will be discussed. Third, an objection will be raised according to which embodied accounts underestimate the cognitive functions language fulfills. The remainder of the paper will be devoted to arguing for (...) the cognitive indispensability of non-embodied, abstract representations by highlighting some of the cognitive benefits they bestow upon us. (shrink)

We review publications in automated scientific discovery using deep learning, with the aim of shedding light on problems with strong connections to philosophy of science, of physics in particular. We show that core issues of philosophy of science, related, notably, to the nature of scientific theories; the nature of unification; and of causation loom large in scientific deep learning. Therefore, advances in deep learning could, and ideally should, have impact on philosophy of science, and vice versa. We suggest lines of (...) further research, and highlight the role ‘theory-based’ AI could have in future developments of the field. (shrink)

In this paper we make the case for the emergence of novel kind of bias with the use of algorithmic decision-making systems. We argue that the distinctive generative process of feature creation, characteristic of machine learning (ML), contorts feature parameters in ways that can lead to emerging feature spaces that encode novel algorithmic bias involving already marginalized groups. We term this bias _assembled bias._ Moreover, assembled biases are distinct from the much-discussed algorithmic bias, both in source (training data versus feature (...) creation) and in content (mimics of extant societal bias versus reconfigured categories). As such, this problem is distinct from issues arising from bias-encoding training feature sets or proxy features. Assembled bias is not epistemically transparent in source or content. Hence, when these ML models are used as a basis for decision-making in social contexts, algorithmic fairness concerns are compounded. (shrink)

The predominance of machine learning based techniques in cognitive neuroscience raises a host of philosophical and methodological concerns. Given the messiness of neural activity, modellers must make choices about how to structure their raw data to make inferences about encoded representations. This leads to a set of standard methodological assumptions about when abstraction is appropriate in neuroscientific practice. Yet, when made uncritically these choices threaten to bias conclusions about phenomena drawn from data. Contact between the practices of multivariate pattern analysis (...) and philosophy of science can help to illuminate the conditions under which we can use artificial neural networks to better understand neural mechanisms. This paper considers a specific technique for MVPA called representational similarity analysis. I develop a theoretically-informed account of RSA that draws on early connectionist research and work on idealization in the philosophy of science. By bringing a philosophical account of cognitive modelling in conversation with RSA, this paper clarifies the practices of neuroscientists and provides a generalizable framework for using artificial neural networks to study neural mechanisms in the brain. (shrink)

The problem of epistemic opacity in Artificial Intelligence is often characterised as a problem of intransparent algorithms that give rise to intransparent models. However, the degrees of transparency of an AI model should not be taken as an absolute measure of the properties of its algorithms but of the model’s degree of intelligibility to human users. Its epistemically relevant elements are to be specified on various levels above and beyond the computational one. In order to elucidate this claim, I first (...) contrast computer models and their claims to algorithm-based universality with cybernetics-style analogue models and their claims to structural isomorphism between elements of model and target system. While analogue models aim at perceptually or conceptually accessible model-target relations, computer models give rise to a specific kind of underdetermination in these relations that needs to be addressed in specific ways. I then undertake a comparison between two contemporary AI approaches that, although related, distinctly align with the above modelling paradigms and represent distinct strategies towards model intelligibility: Deep Neural Networks and Predictive Processing. I conclude that their respective degrees of epistemic transparency primarily depend on the underlying purposes of modelling, not on their computational properties. (shrink)

The topic of this dissertation is the nature of linguistic competence, the capacity to understand and produce sentences of natural language. I defend the empiricist account of linguistic competence embedded in the connectionist cognitive science. This strand of cognitive science has been opposed to the traditional symbolic cognitive science, coupled with transformational-generative grammar, which was committed to nativism due to the view that human cognition, including language capacity, should be construed in terms of symbolic representations and hardwired rules. Similarly, linguistic (...) competence in this framework was regarded as being innate, rule-governed, domain-specific, and fundamentally different from performance, i.e., idiosyncrasies and factors governing linguistic behavior. I analyze state-of-the-art connectionist, deep learning models of natural language processing, most notably large language models, to see what they can tell us about linguistic competence. Deep learning is a statistical technique for the classification of patterns through which artificial intelligence researchers train artificial neural networks containing multiple layers that crunch a gargantuan amount of textual and/or visual data. I argue that these models suggest that linguistic competence should be construed as stochastic, pattern-based, and stemming from domain-general mechanisms. Moreover, I distinguish syntactic from semantic competence, and I show for each the ramifications of the endorsement of a connectionist research program as opposed to the traditional symbolic cognitive science and transformational-generative grammar. I provide a unifying front, consisting of usage-based theories, a construction grammar approach, and an embodied approach to cognition to show that the more multimodal and diverse models are in terms of architectural features and training data, the stronger the case is for the connectionist linguistic competence. I also propose to discard the competence vs. performance distinction as theoretically inferior so that a novel and integrative account of linguistic competence originating in connectionism and empiricism that I propose and defend in the dissertation could be put forward in scientific and philosophical literature. (shrink)