Abstracts of Papers

  List of Accepted Paper Presentations: Abstracts
A1 Aaron Sloman, U Birmingham Abstract: Despite huge practical importance of developments in AI, there have always been researchers (including Turing) less interested in using AI systems to do useful things and more interested in potential of AI as *science* and *philosophy*; in particular the potential to advance knowledge by providing new explanations of natural intelligence and new answers to ancient philosophical questions about what minds are. A particularly deep collection of questions concerns explaining how biological evolution produced so many different forms of intelligence, in humans and non-human animals, and in humans at different stages of development, in different physical and cultural contexts. Current AI still cannot match the amazing mathematical discoveries by ancient mathematicians still in widespread use all over this planet by scientists, mathematicians and engineers. The discovery processes are quite unlike statistical/probabilistic learning for reasons spelled out in Kant's philosophy of mathematics.
A2 Hajo Greif, TU Munich Abstract: Given the personal acquaintance between Alan M. Turing and W. Ross Ashby and the partial proximity of their research fields, a comparative view of Turing’s and Ashby’s work on modelling “the action of the brain” (letter from Turing to Ashby, 1946) will help to shed light on the seemingly strict symbolic/embodied dichotomy: While it is clear that Turing was committed to formal, computational and Ashby to material, analogue methods of modelling, there is no straightforward mapping of these approaches onto symbol-based AI and embodiment-centered views respectively. Instead, it will be demonstrated that both approaches, starting from a formal core, were at least partly concerned with biological and embodied phenomena, albeit in revealingly distinct ways.
A3 Hyungrae Noh, U Iowa Abstract: The receptor notion is that a neural or computational structure has the role of indicating because it is regularly and reliably activated by some distal conditions (Ramsey 2007). The receptor notion has been widely accepted by naturalistic theories of content as a form that combines Shannon-information theory (1948) with teleosemantics (Dretske 1981; Skyrms 2010; Shea et al. 2017).  … cognitive systems may turn out to be as deeply embedded in, hence dependent on, their specific environments as are other biological systems (Millikan 2003). In other words, biological signals (both in the correlational information conveying sense and the mapping functional sense) are affordance like food or like shelter, things that are what they are only relative to an animal who would use them for a specific purpose (Millikan 2012). 
A4 Ioannis Votsis, New College & LSE, London Abstract: What is computation? At the heart of this question appears to lie a paradox. On the one hand, computation looks like the kind of thing virtually any physical system does. After all, physics ensures that some states are followed by other states in a rule-like manner. This view has come to be known as ‘pancomputationailsm’. On the other hand, computation looks like the kind of thing that only emerged in recent human history. On this view, very few physical systems compute, namely those that were technologically designed to do so. We may call this ‘oligocomputationalism’. This talk aims to resolve the apparent paradox by putting forward two non-rivalling notions of computation: one that underwrites pancomputationalism and another that underwrites oligocomputationalism. It is argued that each notion is legitimate because it captures different uses of the term ‘computation’.
A5 Jiri Wiedermann and Jan van Leeuwen, Czech Academy of the Sciences Abstract: AI research is continually challenged to explain cognitive processes as being computational. Whereas existing notions of computing seem to have their limits for it, we contend that the recent, epistemic approach to computations may hold the key to understanding cognition from this perspective. Here, computations are seen as processes generating knowledge over a suitable knowledge domain, within the framework of a suitable knowledge theory. This, machine-independent, understanding of computations allows us to explain a variety of higher cognitive functions such as accountability, self-awareness, introspection, knowledge understanding, free will, creativity, anticipation, curiosity in computational terms, as well as to understand the mechanisms behind the development of intelligence. The argumentation does not depend on any technological analogies.
A6 Matthew Childers, U Iowa Abstract: The “Symbol Grounding Problem” (SGP) concerns how an artificial agent (AA) can autonomously derive the meaning of the symbols it successfully manipulates syntactically. While conditions for a successful solution has been proposed (the “Z/B-condition”), few have considered a teleosemantic solution which meets the Z/B-condition. I argue that a teleosemantic solution is problematic because orthodox teleosemantics construes representation in terms the evolutionary etiology of the biological capacities of representational agents and systems. I assess the strengths of three non-etiological theories of function (propensity, modal, and causal-role theories) and show that they all fail. In turn, I outline avenues for a teleosemantic solution to the SGP afforded by artificial evolution and genetic programming research. Yet these also fall afoul of the Z/B-condition.
A7 Paul Schweizer, U Edinburgh Abstract: The paper explores two related variations on the ‘animat’ theme. Animats are hybrid devices with both artificial and biological components. Traditionally, ‘components’ have been construed in concrete terms, as physical parts or constituent material structures. Many interesting issues arise within this context of hybrid physical organization. However, within the context of functional/computational theories of mentality, demarcations based purely on material structure are far too narrow. It is abstract functional structure which does the key work in characterizing the respective ‘components’ of thinking agents, while the ‘stuff’ of material implementation is of secondary importance. Thus the paper extends the received animat paradigm, and explores some intriguing consequences of expanding the conception of bio-machine hybrids to include abstract functional structure, and not just concrete physical parts. In particular, I extend the animat theme to encompass cases of mental bio-synthetic hybrids.   
A8 Blay Whitby, U Sussex Abstract: A much neglected area in Artificial Intelligence ethics concerns the increasing use of simulated emotional responses. Though these are not ‘genuine’ or in any way equivalent to human emotional responses, there is ample evidence that humans can easily be manipulated by simulated emotional displays made by machines – even when they know it to be simulated and the level of simulation is relatively crude. The technology of artificial emotions is no longer experimental and it is time to analyze what ethical limits should be applied. Since there is no applicable legislation and very close to zero ethical guidance on the ethics of artificial emotions. This is an area which deserves far more attention from those interested in AI ethics.
A9 Al Baker and Simon Wells, Aberdeen U Abstract: Even today, artificial intelligences of varying complexity and sophistication are used for a broad range of persuasive purposes.  My FitBit persuades me to run that extra half mile, while Amazon and Steam persuade me to make new purchases on the basis of previous ones.  These examples are innocuous enough, but it is easy to imagine potentially more troubling uses of persuasive AI.  Should an AI be permitted to persuade a criminal defendant to plead guilty?  Persuade me to pursue a career?  Who to vote for?  Who to date? We discuss the morally relevant differences between persuasion by artificial and human intelligences, how to determine what principles governing human to human persuasion should govern AI persuasion, what additional principles may be necessary, and how those principles can help us decide under what circumstances AI to human persuasion is impermissible.
A10 Geoff Keeling, U Bristol Abstract: Driverless cars will be on our roads soon. Many people argue that driverless cars will sometimes encounter collisions where (i) harm to at least one person is unavoidable or very likely and (ii) a choice about how to allocate harm or expected harm between different persons is required. How should we programme driverless cars to allocate harm in these collisions? Derek Leben proposes a Rawlsian answer to this question. In this paper, I argue that we have good moral reasons to reject Leben’s answer.
A11 Michael Prinzing, U North Carolina Abstract: There is a non-trivial chance that sometime, in perhaps the not too distant future, someone somewhere will build an artificial general intelligence (AI). If that’s true, it seems not unlikely that AI will eventually surpass human-level cognitive proficiency, even may even become “superintelligent”. The advent of superintelligence has great potential—for good or ill. It is therefore imperative that we find a way to guarantee—before one arrives—that any superintelligence we build will remain friendly. Programming an AI to pursue goals that we find congenial will be an extremely difficult challenge. This paper proposes a novel solution to this puzzle: program the AI to love humanity. For friendly superintelligence, I suggest, all you need is love.
A12 Sander Beckers, Cornell U Abstract: The ethical concerns regarding the development of an Artificial Intelligence have received a lot of attention lately. Even if we have good reason to believe that it is very unlikely, the mere possibility of an AI causing extreme human suffering is problematic enough to warrant serious consideration. In this paper I argue that a similar concern arises when we look at this problem from the perspective of the AI. Even if we have good reason to believe that it is very unlikely, the mere possibility of humanity causing extreme suffering to an AI is problematic enough to warrant serious consideration.
A13 Tom Everitt, Australian National University Abstract: How can we maintain control over agents that are smarter than ourselves? We argue that we need to ensure that we build agents that have goals aligned with ours; that are corrigible and won't resist shutdown or corrections; and that strive to preserve these properties in the face of accidents, adversaries, and potential self-modifications.
A14 Torben Swoboda, U Bayreuth Abstract: Sparrow has argued that autonomous weapon systems cause a responsibility gap, because the actions of an AWS are not predictable. In this paper I distinguish between local and non-local behaviour and argue that non-local behaviour is sufficient for attributing responsibility. An AWS can be instantiated by supervised learning. In such a case the programmer is aware of the non-local behaviour of the system. This implies that the programmer is blameworthy and liable for caused damages, whenever the AWS wrongfully causes harm. I then formulate a consequentialist criterion which excuses the programmer from being held responsible. Lastly, I list challenges that remain, so that we should remain sceptical about deploying AWS.
B1 Anna Strasser, Berlin School of Mind & Brain Abstract: Standard notions in philosophy of mind characterize socio-cognitive abilities as if they are unique to sophisticated adult human beings. But soon we are going to be sharing a large part of our lives with various kinds of artificial agents. That is why I will explore in this paper how we can expand these restrictive notions in order to account for other types of agents. Current minimal notions such as minimal mindreading and a minimal sense of commitment present a promising starting point since they show how these notions can be expanded to infants and non-human animals. Considering developments in Artificial Intelligence I will discuss in what sense we can expand our conception of sociality to artificial agents.
B2 Bryony Pierce, U Bristol Abstract: For reasons for action to be grounded, facts concerning those reasons must obtain in virtue of something more fundamental, in a relation of non-causally dependent justification. I argue that, in a robot or other entity with artificial intelligence, grounding would have to be in something external: the qualitative character of the affective responses of its programmers or other human beings.  I explore a number of senses of grounding and discuss the distinction between semantic and affective content in the context of grounding reasons for action.
B3 Chuanfei Chin, National University of Singapore Abstract: What new challenges are emerging in philosophical debates on artificial consciousness? Earlier debates focused on thought-experiments such as Block’s Chinese Nation and Searle’s Chinese Room, which seemed to show that phenomenal consciousness cannot arise, or cannot be produced, in non-biological machines. These debates on the possibility of artificial consciousness have been transformed as we make more use of empirical methods, models, and evidence. I shall argue that this naturalistic approach leads to a new set of philosophical challenges. These challenges centre on the multiplicity of neurofunctional structures which underlie ‘what it is like’ to be in a conscious state. When we uncover more than one kind of phenomenal consciousness, how should we conceptualise artificial consciousness? By addressing this challenge, we can classify different theories of artificial consciousness in the AI literature and clarify the moral status of any conscious machines.
B4 David Longinotti, University of Maryland Abstract: Non-living machines can’t be agents, nor can they be conscious. An action is homeostatic for its agent and must begin in something that moves to maintain itself: living matter. Behavior motivated by feelings is homeostatic, so the source of qualia is a living substance. This can also be inferred scientifically. Various empirical evidence indicates that feelings are a distinct form of energy generated in specialized neurons. Phenomenal energy would not be objectively observable if it were spent as and where it is produced. This expenditure is thermodynamically necessary if, by converting action potentials to feelings, the source of the feelings averts an increase in its entropy. Thus, the source of qualia is a living, self-maintaining substance.
B5 Rene Mogensen, U Birmingham City Abstract: Geraint A. Wiggins proposed a formalised framework for ‘computational creativity’, based on Margaret Boden’s view of ‘creativity’ defined as searches in conceptual spaces. I argue that the epistemological basis for well-defined ‘conceptual spaces’ is problematic: instead of Wiggins’s well-defined types or sets, such theoretical spaces can represent emergent traces of creative activity. To address this problem, I have revised the framework to include dynamic conceptual spaces, along with formalisations of memory and motivations, which allow iteration in a time-based framework that can be aligned with experiential learning models (e.g., John Dewey’s). My critical revision of the framework, applied to the special case of improvising computer systems, achieves a more detailed specification and better understanding of computational creativity.
B6 Sankalp Bhatnagar, Shahar Avin, Stephen Cave, Marta Halina, Aiden Loe, Seán Ó HÉigeartaigh, Huw Price, Henry Shevlin and Jose Hernandez-Orallo, U of Cambridge Abstract: New types of artificial intelligence (AI), from cognitive assistants to social robots, are challenging meaningful comparison with other kinds of intelligence. How can such intelligent systems be catalogued, evaluated, and contrasted, so that representations and projections offer more meaningful insights? AI and the future of cognition research can be catalyzed by an alternative framework and collaborative open repository for collecting and exhibiting information of all kinds of intelligence, including humans, non-human animals, AI systems, hybrids and collectives thereof. After presenting this initiative, we review related efforts and offer the results of a pilot survey on the motivations, applications and dimensions of such a framework, aimed at identifying and refining its requirements and possibilities
B7 Yoshihiro Maruyama, U of Oxford Abstract: The frame problem is a fundamental challenge in AI, and the Lucas-Penrose argument indicates a limitation of AI if it is successful. We discuss both of them from a unified Gödelian point of view. We give an informational reformulation of the frame problem, which turns out to be tightly linked with the nature of Gödelian incompleteness. We then revisit the Lucas-Penrose argument, giving a version of it which shows the impossibility of information physics. It then turns out through a finer analysis that if the Lucas-Penrose argument is accepted then information physics is impossible too. Our arguments indicate that the frame problem and the Lucas-Penrose argument share a common Gödelian structure at a level of abstraction, and what is crucial for both is the finitarity condition of frame and computation, without which the limitations can readily be overcome. 
B8 Abhishek Mishra, National University of Singapore Abstract: While there has been much discussion about the moral status of humans, non-human animals and even other natural entities, discussion about the moral status of digital agents has been limited. This paper proposes a way of reasoning about how we should act towards digital agents under moral uncertainty by considering the particular case of how we should act towards simulations run by an artificial superintelligence (ASI). By placing the problem of simulations within the larger problem of AI-safety (how to ensure a desirable post-ASI outcome) as well as debates about the grounds of moral status, this paper formalizes it into a decision problem. The paper ends by suggesting future steps to solve this decision problem, and how the approach might be generalized.
B9 Ron Chrisley, U Sussex Abstract: I argue that for auto-epistemic knowledge-based systems, a further constraint beyond consistency, which I call *epistemic consistency*, must be met.  I distinguish two versions of the constraint: *propositional* epistemic consistency requires that there be no sentences in an agent’s knowledge base that constitute an epistemic blindspot for that agent. Maintaining this constraint requires generalising from the notion of an epistemic blindspot to the concept of epistemic blindspot sets; I show how enforcing this requirement can prevent fallacious reasoning of a form found in some well-known paradoxes. The other version, *inferential* epistemic consistency, forbids certain epistemically problematic inferences.  I argue that the intuitive notion of the validity of a rule of inference can only be retained if inferential epistemic consistency is enforced.
B10 Daniel Kokotajlo and Ramana Kumar, U North Carolina Abstract: We articulate two projects that decision theorists can engage in: Roughly, they are (a) trying to discover the norms that govern instrumental reasoning, and (b) trying to decide which decision procedures to install in our AIs. We are agnostic about the relationship between the two projects, but we argue that the two most popular answers to (a), CDT and EDT, are clearly not good answers to (b). Our overall goal is to argue that project (b) exists, that it is immensely important, and that decision theorists can productively contribute to it. Indeed, perhaps some decision theorists already are; this is what we take the Machine Intelligence Research Institute to be doing.
B11 J Mark Bishop, John Howroyd and Andrew Martin, Goldsmiths, U London Abstract: In this paper we demonstrate [for the first time] progress towards developing a swarm intelligence algorithm - based on interacting communicating processes - that is Turing complete. This is a relatively an important result – not least as the core principle underlying the interacting processes are (a) analogous to the behaviour of certain [tandem running] ant species (in nest/resource location tasks) and (b) are based on communications NOT computations (although, of course, they can be described [and simulated] computationally); the latter feature, positions out work in a different class to both Siegelmann and Sontag’s Turing Complete RNN (Recurrent Neural Network), and the Google/DeepMind team (of Grave, Wayne &amp; Danihelka) 2014 NTM (Neural Turing Machine), both of which remain implicitly – if not explicitly - grounded on computational processes (summations, multiplications, activation-functions etc.).</body>
B12 Raül Fabra Boluda, Cesar Ferri, Jose Hernandez-Orallo, Fernando Martínez-Plumed and M.J. Ramírez, U Sevilla Abstract: Being surrounded by machine learning (ML) models making decisions for governments, companies and individuals, there is the increasing concern of not having a rich explanatory and predictive account of the behaviour of these ML models relative to the users' interests (goals) and (pre-)conceptions (narratives). We argue that the recent research trends in finding better characterisations of what a ML model does are leading to the view of ML models as complex behavioural systems. Consequently, we argue that a more contextual abstraction is necessary, as done in system theory and psychology, very much like a mind modelling problem. We bring some research evidence of how this transition can take place, suggesting that more machine learning is the answer.
B13 Shlomo Danziger, Hebrew U of Jerusalem Abstract: Turing's Imitation Game (IG) is usually understood as a test for machines' intelligence. I offer an alternative interpretation, according to which Turing holds an externalist-like view of intelligence; and I discuss some ramifications this view may have on current AI development and cognitive research. Turing, I argue, conditioned the determination that a machine is intelligent upon two criteria: one technological and one sociolinguistic. The technological criterion, tested by the IG, requires that the machine be designed so that its behavior is indistinguishable from human intellectual behavior. But the IG does not test if the machine is intelligent; that requires also the fulfillment of the sociolinguistic criterion – that the machine be perceived by society as a potentially intelligent entity. To Turing, intelligence is constituted by the way a system is perceived by humans, and not just by its internal properties.
B14 Tobias Wängberg, Mikael Böörs, Elliot Catt, Tom Everitt and Marcus Hutter, Australian National University Abstract: The off-switch game is a game theoretic model of a highly intelligent robot interacting with a human. In the original paper by Hadfield-Menell et al. (2016), the analysis is not fully game-theoretic as the human is modelled as an irrational player, and the robot's best action is only calculated under unrealistic normality and soft-max assumptions. Wangberg et al. (2017) make the analysis fully game theoretic, by modelling the human as a rational player with a random utility function. As a consequence, the robot's best action can be calculated for arbitrary belief and irrationality assumptions.