Michael Anderson

At the turn of the 21st century, Susan Leigh Anderson and Michael Anderson conceived and introduced the Machine Ethics research program, that aimed to highlight the requirements under which autonomous artificial intelligence systems could demonstrate ethical behavior guided by moral values, and at the same time to show that these values, as well as ethics in general, can be representable and computable. Today, the interaction between humans and AI entities is already part of our everyday lives; in the near future it is expected to play a key role in scientific research, medical practice, public administration, education and other fields of civic life. In view of this, the debate over the ethical behavior of machines is more crucial than ever and the search for answers, directions and regulations is imperative at an academic, institutional as well as at a technical level. Our discussion with the two inspirers and originators of Machine Ethics highlights the epistemological, metaphysical and ethical questions arising by this project, as well as the realistic and pragmatic demands that dominate artificial intelligence and robotics research programs. Most of all, however, it sheds light upon the contribution of Susan and Michael Anderson regarding the introduction and undertaking of a main objective related to the creation of ethical autonomous agents, that will not be based on the “imperfect” patterns of human behavior, or on preloaded hierarchical laws and human-centric values.

The current paper details a restricted semantics for active logic, a time-sensitive, contradictiontolerant logical reasoning formalism. Central to active logic are special rules controlling the inheritance of beliefs in general, and beliefs about the current time in particular, very tight controls on what can be derived from direct contradictions (P &¬P ), and mechanisms allowing an agent to represent and reason about its own beliefs and past reasoning. Using these ideas, we introduce a new definition of model and of logical consequence, as well as a new definition of soundness such that, when reasoning with consistent premises, all classically sound rules are sound for active logic. However, not everything that is classically sound remains sound in our sense, for by classical definitions, all rules with contradictory premises are vacuously sound, whereas in active logic not everything follows from a contradiction.

Many commentators on Clark’s writings on predictive processing have wondered how well the predictive processing model actually fits with embodied and extended cognition. The former seems to imply that cognition is secluded from the environment, while the latter implies that cognition is in and of the environment. This chapter argues that a reconciliation with embodied and extended cognition is possible but requires that predictive processing proponents reject environmental seclusion. To do so means adopting ecological information in place of the Shannon information most typically invoked by proponents of predictive processing, and giving many of the other semantic-sounding terms they use (e.g., “prediction,” “model,” “representation”) deflationary understandings.

To accept that cognition is embodied is to question many of the beliefs traditionally held by cognitive scientists. One key question regards the localization of cognitive faculties. Here we argue that for cognition to be embodied and sometimes embedded, means that the cognitive faculty cannot be localized in a brain area alone. We review recent research on neural reuse, the 1/f structure of human activity, tool use, group cognition, and social coordination dynamics that we believe demonstrates how the boundary between the different areas of the brain, the brain and body, and the body and environment is not only blurred but indeterminate. In turn, we propose that cognition is supported by a nested structure of task-specific synergies, which are softly assembled from a variety of neural, bodily, and environmental components (including other individuals), and exhibit interaction dominant dynamics.

In this essay we respond to some criticisms of the guidance theory of representation offered by Tom Roberts. We argue that although Roberts’ criticisms miss their mark, he raises the important issue of the relationship between affordances and the action-oriented representations proposed by the guidance theory. Affordances play a prominent role in the anti-representationalist accounts offered by theorists of embodied cognition and ecological psychology, and the guidance theory is motivated in part by a desire to respond to the critiques of representationalism offered in such accounts, without giving up entirely on the idea that representations are an important part of the cognitive economy of many animals. Thus, explorations of whether and how such accounts can in fact be related and reconciled potentially offer to shed some light on this ongoing controversy. Although the current essay hardly settles the larger debate, it does suggest that there may be more possibility for agreement than is often supposed.

“The physics of representation” aims to define the word “representation” as used in the neurosciences, argue that such representations as described in neuroscience are related to and usefully illuminated by the representations generated by modern neural networks, and establish that these entities are “representations in good standing”. We suggest that Poldrack succeeds in, exposes some tensions between the broad use of the term in neuroscience and the narrower class of entities that he identifies in the end, and between the meaning of “representation” in neuroscience and in psychology in, and fails in. This results in some hard choices: give up on the broad scope of the term in neuroscience or continue to embrace the broad, psychologically inflected sense of the term, and deny the entities generated by neural nets are representations in the relevant sense.

Cognitive scientists and neuroscientists typically understand the brain as a complex communication/information-processing system. A limitation of this framework is that it requires cognitive systems to have prior knowledge about their environment to successfully perform some of their basic functions, such as perceiving. It is unclear how the source of such knowledge can be explained from within this framework. Drawing on Dennett (1981), we refer to this as the loans of intelligence problem. Recent advances in machine learning have resulted in the development of a family of algorithms, including the class known as autoencoders, that seem to provide a way for the information-processing framework to avoid this problem: cognitive systems do not require loans of intelligence, but instead acquire the knowledge necessary for perception through a process of unsupervised learning. This paper argues that although autoencoders do avoid the loans of intelligence problem, how they do so should not be understood from within the information-processing framework. Instead, their success should be interpreted as a proof of concept of how neural networks can attune to Gibsonian information. We thus propose that autoencoders belong to a class of algorithms for modeling the brain that have recently been dubbed direct fit algorithms.

Within the ecological and enactive approaches in cognitive science, a tension exists in how the process of skill learning is understood. Skill learning can be understood in a narrow sense, as a process of bodily change over time, or in an extended sense, as a change in the structure of the animal–environment system. We propose to resolve this tension by rejecting the first understanding in favor of the second. We thus defend an extended approach to skill learning. An extended understanding of skill learning views bodily changes as being embedded in a larger process of interaction between the organism and specific structures in the environment. Such an extended approach is committed to the claims that (1) the appropriate unit of analysis for understanding skill learning is not the body but the activity and (2) learning consists in the establishment and adaptive organization of enabling constraints on that activity. We focus on two example cases: maintaining upright posture and walking. In both cases, environmental structures play a constitutive role in the activity throughout learning, but the specific environmental structures that are involved in the activity change over time. At an early stage, the child makes use of an environmental “support”—for example, holding onto furniture to maintain upright posture. Later, once further constraints have been established, the child is able to let go of the furniture and remain upright. We argue that adopting an extended understanding of skill learning offers a promising strategy for unifying ecological and enactive approaches and can also potentially ground a radically embodied approach to higher cognition.

To think about how to anchor abstract symbols to objects in the world is to become part of a tradition in philosophy with a long history, and an especially rich recent past. It is to ask, with Wittgenstein, “What makes my thought about him, a thought about him?” and thus it is to wonder not just about the nature of referring expressions or singular terms, but about the nature of referring beings. With this in mind I hereby endeavor—briefly, incompletely, but hopefully still usefully—to introduce what in my judgment is the single best philosophical starting-point for those interested in understanding the referential connections between symbols and the world, and the cognitive, epistemic, and linguistic capacities which support them: The Varieties of Reference by Gareth Evans.1 It is worthwhile first of all to note, as the title indicates, that it is the varieties of reference that are of interest. It is Evans’ contention that no single theory can account for our various use of singular terms; although the different kinds of reference share certain features, and rely on related cognitive, linguistic and epistemic capacities, it appears that, rather than being a class defined by necessary and sufficient criteria for membership, they form a family of abilities, united, like a thread, by its overlapping fibers. Evans does not defend this claim so much as display it in his account. Much of the underlying variety in reference can be brought out by considering the guiding principle of the work as a whole, which Evans..

One of the most foundational and continually contested questions in the cognitive sciences is the degree to which the functional organization of the brain can be understood as modular. In its classic formulation, a module was defined as a cognitive sub-system with nine specific properties; the classic module is, among other things, domain specific, encapsulated, and implemented in dedicated neural substrates. Most of the examinations—and especially the criticisms—of the modularity thesis have focused on these properties individually, for instance by finding counterexamples in which otherwise good candidates for cognitive modules are shown to lack domain specificity or encapsulation. The current paper goes beyond the usual approach by asking what some of the broad architectural implications of the modularity thesis might be, and attempting to test for these. The evidence does not favor a modular architecture for the cortex. Moreover, the evidence suggests that best way to approach the understanding of cognition is not by analyzing and modelling different functional domains in isolation from the others, but rather by looking for points of overlap in their neural implementations, and exploiting these to guide the analysis and decomposition of the functions in question. This has significant implications for the question of how to approach the design and implementation of intelligent artifacts in general, and language-using robots in particular.

Recent trends in the philosophy of mind and cognitive science can be fruitfully characterized as part of the ongoing attempt to come to grips with the very idea of homo sapiens--an intelligent, evolved, biological agent--and its signature contribution is the emergence of a philosophical anthropology which, contra Descartes and his thinking thing, instead puts doing at the center of human being. Applying this agency-oriented line of thinking to the problem of representation, this paper introduces the Guidance Theory, according to which the content and intentionality of representations can be accounted for in terms of the way they provide guidance for action. We offer a brief account of the motivation for the theory, and a formal characterization

As is well known, dialog partners manage the uncertainty inherent in conversation by continually providing and eliciting feedback, monitoring their own comprehension and the apparent comprehension of their dialog partner, and initiating repairs as needed (see e.g., Cahn & Brennan, 1999; Clark & Brennan, 1991). Given the nature of such monitoring and repair, one might reasonably hypothesize that a good portion of the utterances involved in dialog management employ meta-language. But while there has been a great deal of work on the specific topic of dialog management, and it is widely (if often tacitly) accepted that meta-language is frequently involved, there has been no work specifically investigating and quantifying the role of meta-language in dialog management. Thus, this small study investigated the correlation between meta-language and dialog management utterances in three dialog files of the British National Corpus (BNC).

This paper lays out some of the empirical evidence for the importance of neural reuse—the reuse of existing (inherited and/or early-developing) neural circuitry for multiple behavioral purposes—in defining the overall functional structure of the brain. We then discuss in some detail one particular instance of such reuse: the involvement of a local neural circuit in finger awareness, number representation, and other diverse functions. Finally, we consider whether and how the notion of a developmental homology can help us understand the relationships between the cognitive functions that develop out of shared neural supports

However, there has also been growing interest in trying to create, and investigate the potential benefits of, intelligent systems which are themselves metacognitive. It is thought that systems that monitor themselves, and proactively respond to problems, can perform better, for longer, with less need for (expensive) human intervention. Thus has IBM widely publicized their "autonomic computing" initiative, aimed at developing computers which are (in their words) self-aware, selfconfiguring, self-optimizing, self-healing, self-protecting, and self-adapting. More ambitiously, it is hypothesized that metacognitive awareness may be one of the keys to developing truly intelligent artificial systems. DARPA's recent Cognitive Information Processing Technology initiative, for instance, foregrounds reflection (along with reaction and deliberation) as one of the three pillars required for flexible, robust AI systems.

Multi-voxel pattern analysis (MVPA) is a popular analytical technique in neuroscience that involves identifying patterns in fMRI BOLD signal data that are predictive of task conditions. But the technique is also frequently used to make inferences about the regions of the brain that are most important to the tasks in question, and our analysis shows that this is a mistake. MVPA does not provide a reliable guide to what information is being used by the brain during cognitive tasks, nor where that information is. This is due in part to inherent run to run variability in the decision space generated by the classifier, but there are also several other issues, discussed below, that make inference from the characteristics of the learned models to relevant brain activity deeply problematic. These issues have significant implications both for many papers already published, and for how the field uses this technique in the future.

Maintaining adequate performance in dynamic and uncertain settings has been a perennial stumbling block for intelligent systems. Nevertheless, any system intended for real-world deployment must be able to accommodate unexpected change—that is, it must be perturbation tolerant. We have found that metacognitive monitoring and control—the ability of a system to self-monitor its own decision-making processes and ongoing performance, and to make targeted changes to its beliefs and action-determining components—can play an important role in helping intelligent systems cope with the perturbations that are the inevitable result of real-world deployment. In this article we present the results of several experiments demonstrating the efficacy of metacognition in improving the perturbation tolerance of reinforcement learners, and discuss a general theory of metacognitive monitoring and control, in a form we call the metacognitive loop.

The current essay introduces the guidance theory of representation, according to which the content and intentionality of representations can be accounted for in terms of the way they provide guidance for action. The guidance theory offers a way of fixing representational content that gives the causal and evolutionary history of the subject only an indirect role, and an account of representational error, based on failure of action, that does not rely on any such notions as proper functions, ideal conditions, or normal circumstances. Moreover, because the notion of error is defined in terms of failure of action, the guidance theory meets the “meta-epistemological requirement” that representational error should be potentially detectable by the representing system itself. In this essay, we offer a brief account of the biological origins of representation, a formal characterization of the guidance theory, some examples of its use, and show how the guidance theory handles some traditional problem cases for representation: the representation of fictional and abstract entities. Being both representational and actiongrounded, the guidance theory may provide some common ground between embodied and cognitivist approaches to the study of the mind