Modeling Minds II: Abstracts

Speaker: Marc Slors

Title: Levels, perspectives, interpretations

In most theorizing about different levels of description in cognitive science, inter-level relations are thought to be objective or observer-independent. A physical (i.e. ‘low’ level) description of computer hardware, for instance, entails its functioning (i.e. a ‘higher’ level description), regardless of what we think it does. Similarly, lower-level descriptions of cognitive systems (say in terms of Marr’s algorithmic level) are treated as conceptually entailing higher-level descriptions (say, in terms of Marr’s computational level). This does not apply, however, to the relation between the levels of description that follow from adopting what Daniel Dennett calls the intentional, design and physical stance. Although the patterns that characterize the intentional level of description are objective, according to Dennett, we need an interpretive strategy or stance to discern them. That is, they are not simply given, regardless of who is doing the interpreting. In this talk I will argue (1) that Dennett’s approach precludes a tractable implementation or realization relation between levels, (2) that this allows for the idea that the sub-intentional levels of description of intentional systems need not be recognizably representational, intentional or mental, and (3) that, unlike traditional views, this approach can therefore reconcile an enactive, non-representational conception of cognition at the level of biological design with an intentional, higher-level description in terms of e.g. beliefs and desires.

Speaker: Bart Geurts

Title: No sense without competence

David Marr famously characterised his three levels in terms of the signature questions associated with them. I believe that speaking and/or thinking in terms of levels can be quite misleading, and prefer to focus on the questions instead of the levels. In particular, I will consider the question what a cognitive system does or can do: its competence. In this talk, I will present two case studies based on my own work in (theoretical and experimental) pragmatics: one involves the interpretation of conversational implicatures, the other concerns the production of referential expressions. With the help of these examples, I will defend the following claims:

Speaker: Willem Frankenhuis

Title: Levels of explanation in biology

Nobel-Prize winning ethologist Niko Tinbergen argued that a complete explanation of any given trait or behavior involves four different levels of explanation: causation, development, function, and phylogeny. Most psychologists adopt a proximate perspective, focusing on causation (neural, cognitive, or physiological processes) and development (changes occurring over the lifespan of individuals). Few adopt an ultimate perspective, focusing on phylogeny (evolutionary history of traits) and function (adaptive value of traits). I will argue that these levels of explanation are mutually informative, and show how ideas about adaptive function can inform studies of sensitive periods in development—periods in which experience shapes traits or behaviors to a larger extent than other periods. I use mathematical modeling to better understand how natural selection shapes development to respond to environmental conditions. The goal of this research is to explain adaptive variation in sensitive periods (1) between species in the same trait (e.g., bird song), (2) between individuals within populations (e.g., differential susceptibility), and (3) between traits within a single individual (e.g., differential adjustment of cognitive vs. emotional systems following adoption).

Frankenhuis, W. E., Panchanathan, K., & Barrett, H. C. (2013). Bridging developmental systems theory and evolutionary psychology using dynamic optimization. Developmental Science, 16, 584-598.

Speaker: Pim Haselager

Title: Levels of agency

Debates about agency and the sense of agency generally encompass several levels of analysis. Characterizations of what an act is, what is required to consider a system as an agent, and how systems can appreciate the difference between an event and an act, normally invoke concepts or data from Marr’s computational, algorithmic and implementation levels. However, often finer-grained distinctions are being made, such as in Pacherie’s ‘cascade model’, or different types of levels are invoked, e.g. in typical ‘action hierarchies’. Based upon several examples I will suggest that part of the complexity of the issue of agency stems from an unsystematic and unconstrained use of the notion of levels.





Modeling Minds: Abstracts

Speaker: Catarina Dutilh Novaes

Title: Reasoning biases and non-monotonic logics: the case of preferential logics

(joint work with Herman Veluwenkamp)

Non-monotonic logics form a well-established group of theories both in philosophy and in artificial intelligence/computer science, but which have for the most part been neglected by psychologists and cognitive scientists working on reasoning. The main exception is the pioneering work of Stenning and van Lambalgen (2008; 2010). In this paper, we examine a group of experimental results not addressed by Stenning and van Lambalgen, namely the belief bias data, from the point of view of non-monotonic logics. Moreover, instead of adopting their preferred non-monotonic framework, closed-world reasoning, we adopt the family of non-monotonic logics known as preferential logics as the formal background for the discussion. The application of the framework to these empirical results, if successful, would suggest that these logics may represent a plausible descriptive model of human reasoning. However, the comparison with the data will also highlight the limitations of this framework. Indeed, while many instances in which participants seem to be performing deductive reasoning incorrectly can be equally explained as instances of participants in fact correctly performing defeasible reasoning, some of the experimental results to be discussed cannot be straightforwardly explained from the point of view of preferential logics. Instead, we argue that the data that preferential logics cannot account for can be more fruitfully analyzed from the point of view of belief-revision theory, in particular with the concept of screened revision. We conclude that, while it offers valuable insights into the nature of human reasoning, preferential logics are ultimately inadequate as formal models of the phenomena in question. Finally, these results provide the background for a number of general remarks on the very idea of 'modeling minds' with formal tools.

Speaker: Matteo Colombo

Title: Bayesian cognitive science, under-considered alternatives, and the value of specialization

(Joint work with Rogier de Langhe - UGhent)

Bayesian decision theory is a modelling framework ever more prominent in the cognitive and brain sciences. Within this framework, the uncertainty of an agent facing some task is represented with probability distributions, which get dynamically updated in accordance with the Bayesian rule of conditionalization as the agent receives new information. A widely held belief in the cognitive and brain sciences is that the Bayesian framework should be chosen for explaining cognitive phenomena whose production involves uncertainty. However, this belief is far from being unproblematic. There are several under-considered alternatives to Bayesian decision theory as frameworks for representing and dealing with uncertainty, and it is controversial that the Bayesian framework enjoys a special explanatory status over available alternatives. In this talk, I focus on the explanatory role of modelling frameworks in cognitive science. Specifically, I formulate and assess two general arguments for the adoption of Bayesian decision theory as an explanatory framework in cognitive science: the argument from uncertainty, and the argument from specialization. In doing so, I hope to contribute to the overall theme of the workshop by showing how the explanatory value of a modelling framework does not lie only in the intrinsic properties of the framework itself, but it also depends on wider, social considerations concerning the optimal distribution of cognitive labour in scientific practice. [A preprint of the full paper can be found here: ]

Speaker: Sanneke de Haan

Title: Does explanation require the uncovering of underlying mechanisms?

Traditionally, causal explanations are distinguished from mere descriptions of phenomena. In psychiatry for instance, the manifest symptoms of a disorder are distinguished from their underlying causes. And this makes sense: by merely describing the depressed state, the symptoms, one has after all not yet revealed what caused this state. This division testifies of a so called ‘latent variable model’ in which the descriptions refer to manifest variables (e.g. symptoms), and causal explanation involves revealing the latent variable(s) (e.g. underlying mechanism(s)). The latent variable model, however, is problematic in several respects. Firstly, the assumption that each manifest variable has only one latent (causing) variable is question begging. Secondly, this model does not allow for causal influences between manifest variables (cf. Borsboom and Cramer 2013). But the most fundamental problem is that the dichotomy between latent and manifest variables induces a dualistic view on the relation between the descriptive level and causal level, where the descriptive level refers to experiences/phenomenology and the causal level refers to physiological processes. This dualism in turn invites a reductionist solution to the status of their relation. There are alternative models such as dynamical systems theory and network models that do allow for multiple causality, feedback loops, and non-linear causal influences. These seem to be better apt at capturing the complexities of real life phenomena. But if one adopts these models, how can one conceive of what counts as an explanation or a mechanism? If a mechanism does not refer to a latent variable, then what does it refer to? I would like to explore the idea that dynamical systems model patterns and that the description of the pattern itself may be all there is to it. That is, there need not be anything underlying these patterns that is doing the causal work, and hence describing the pattern may be as explanatory as one can get. Taking this alternative picture seriously also invites a re-thinking of the relation between experiences and physiological processes in mereological rather than reductionist terms. I will illustrate this by coming back to psychiatric disorders.

Speaker: Johan Kwisthout

Title: A complexity-theoretic perspective on approximate Bayesian inferences

In the last few decades, many probabilistic or Bayesian computational cognitive models have been proposed as alternative to symbolic or sub-symbolic models. However, even approximate Bayesian computations are known to be intractable in general for real-world models (in contrast to toy examples). This intractability is problematic: if the model does not scale to real-world situations, its purpose in explaining human cognitive capacities is serverly limited. Can we have computational models that are both Bayesian and tractable? In this talk I will discuss parameterized complexity analysis as an indispensable tool for the cognitive modeler, and show that the answer to the previous question is a qualified 'yes'. Using parameterized complexity analysis we can identify situational constraints that (when met in reality) allow for tractable computations, even when the model is scaled to real-world situations.

Speaker: Pim Haselager

Title: Robots as models

Robots provide an interesting platform for modeling cognition and behaviour. Various forms of robot modeling approaches exist, ranging from software bots and virtual agents to physically instantiated humanoids. Starting with Bechtel’s (2008) analysis of the role of mechanisms in the explanation of cognitive phenomena, I will investigate the methodological requirements for a proper assessment of the explanatory value of robot models. In my presentation I will focus on how theory and technology can, and perhaps even should, be combined in a research cycle in order to provide scientifically relevant robot models of mind.

Speaker: Willem Zuidema

Title: Nonsymbolic models of compositional semantics

Arguments for symbolic accounts of cognition, or more specifically a symbolic ‘language of thought’, often are arguments from personal incredulity: symbolists can’t imagine any way that (high-level) cognition could operate other than based on rules & variables. This is perhaps most strongly felt in the domain of natural language, where formal grammars and logics have proven very useful for accounting for grammaticality judgments, semantic inferences and other aspects of language processing. In my talk, I will discuss my recent work with Phong Le (ILLC, UvA) on developing a model of perhaps the most important feature of natural language: the fact that the meanings of sentences are (more or less) systematically derived from the meanings of words and the way they are put together (‘compositional semantics’). Our model, called ‘Inside-Outside Semantics’ (Le & Zuidema, 2014), is based on the Recursive Neural Network architecture (Goller & K¨uchler, 1996; Socher et al., 2010) and obtains state-ofthe-art results in various benchmark tests in computational linguistics. I will show that such models undermine the incredulity about non-symbolic accounts of compositional semantics. Finally, in the spirit of Box’s ‘all models are wrong, but some models are useful’, I will argue that the symbolic-nonsymbolic dichotomy is unproductive, and that both types of models can be complementary at different levels of description.

Speaker: Ralf Cox

Title: Fractal coordination dynamics in human behaviour

It is becoming increasingly evident that temporal variability in behaviour constitutes a rich source of information about the coordinative basis of that behaviour. In this presentation I will emphasize the fractal nature of coordination dynamics underlying human performance in the motor domain and beyond. Fractal scaling is ubiquitous throughout the motor and cognitive systems, but its nature and implications are still subject of speculation. And although the empirical record clearly underlines the importance of fractal scaling, the lack of concrete (explanatory) models prevents the field to fully deliver what it promises. A general framework for understanding fractality in terms of complex dynamical systems and self-organized criticality will be given, and its implications for coordination dynamics will be discussed. Research examples will be given to demonstrate that temporal patterns of variability in behaviour are non-random and that fractal scaling appears more clearly for higher skill levels and with learning.

Speaker: Nina Gierasimczuk

Title: Investigating human reasoning through the lens of logic and games

This talk will concern psychological relevance of logical models for deductive reasoning. I will present an analysis of logical reasoning in a deductive version of the Mastermind game implemented within a popular Dutch online educational learning system (Math Garden). This approach, based on the semantic tableaux method (known from proof-theory), allows deriving predictions about the empirical difficulty of reasoning items. I will also discuss other recent attempts at employing mathematical logic tools to predict and explain the behavior of human subjects engaged in problem solving.