Neurocomputational Models Of Cognitive Development And Processing - Proceedings Of The 14th Neural Computation And Psychology Workshop

Proceedings of the 14th Neural Computation and Psychology Workshop

Katherine Twomey, Alastair Smith;Gert Westermann;Padraic Monaghan;

• Publisher: WSPC
• Year: 2016
• ISBN:
  9789814699358
• Topic: Biological Sciences
• Subtopic: Science General
• Index: Biological Sciences

This volume presents peer-reviewed versions of papers presented at the 14th Neural Computation and Psychology Workshop (NCPW14), which took place in July 2014 at Lancaster University, UK. The workshop draws international attendees from the cutting edg

How to Design Emergent Models of Cognition for Application-Driven Artificial Agents

Serge Thill^* and David Vernon^†

Interaction Lab,
School of Informatics,
University of Skövde,
541 28 Skövde, Sweden
E-mail: ^* [email protected], ^†[email protected]

Emergent models of cognition are attractive for artificial cognitive agents because they overcome the brittleness of systems that are fully specified in axiomatic terms at design time, increasing, for example, the ability to deal with uncertainty and unforeseen events. When the agent is created to fulfil specific requirements defined by a given application, there is an apparent conflict between the emergent (i.e. self-defining) nature of the agent’s behaviour and the pre-specified (i.e. axiomatically-defined) nature of the requirements.

Here, we develop a framework for the design of emergent models of cognition whose behaviour can be shaped to fulfil application requirements while retaining the desired characteristics of emergence. We achieve this by viewing the artificial agent as forming an eco-system with the environment in which it is deployed. Consequently, the objective function that determines the agent’s behaviour is cast in terms that factor in interaction with the environment (while not being controlled by it) and therefore implicitly includes the application requirements.

This framework is particularly relevant to application driven research where artificial agents are designed to interact with humans in a certain manner. We illustrate this with the example of robot-enhanced therapy for children with autism spectrum disorder.

Keywords: Emergent models of cognition; Application-driven research; Robot-enhanced therapy.

1.Models in cognitive science and new challenges from embodied theories

Typically, models of cognition serve the study of cognition itself. McClelland put it rather succinctly when he argued¹ that we should think of models as “tools for exploring the implications of ideas” (p 12), even though they are necessarily simplifications of the real thing. Traditional cognitive science is replete with examples from many paradigms (whether symbolic, subsymbolic, Bayesian, or others) that have successfully done just that.

Models have been useful in the cognitive sciences because they tend to be explanatory models of cognitive mechanisms. With the advent of embodied theories of cognition (according to which human cognition cannot simply be entirely reduced to abstract, amodal symbol manipulation), however, new challenges have appeared. These theories often lead, for instance, to a need for “embodying” the model. This is typically done with an artificial agent, which can be simulated or real.

These approaches have significant hurdles to overcome if the aim is to provide an explanatory model of human cognition. Robotic models, for instance, are limited by the robotic hardware available, in particular the fact that these do not provide anything resembling a human body, even if the robots used are described as “humanoid” and “embodied”. Models “embodied” in simulated environments, meanwhile, lose a significant degree of realism since simulators simply cannot approximate the complex physics of reality. Like robotic models, they raise the question of whether a (simulated) non-human body is an acceptable abstraction of a human body but extend the issue to the environment itself.

It has been argued repeatedly^2,3 that an advantage of robotic models of cognition is that they are forcibly integrated because every process from sensory perception to the mechanisms of the cognitive behaviours of interest needs to be modelled. This is rarely true, however, as readily illustrated by – for instance – the complexities of computer vision forcing modellers to take short cuts in obtaining visual inputs, for example by assuming a process which returns coordinates of objects of interest⁴ or by using coloured, easily-discriminable objects⁵. The promised “complete” modelling, from sensory perception to higher-level cognition is thus not necessarily given and adds to the non-human embodiment used in simplified environments (even when real robots are used) to create a model embodied in something that is entirely different from a human experience. In particular, although sensorimotor aspects are thought to fundamentally shape higher cognition, they are often the first to be simplified.

Additionally, it is not a given (even though this is often assumed) that merely instantiating a model in a robotic body overcomes the limitations of traditional symbolic approaches to (strong) AI that such models typically intend to address (by virtue of taking an embodied, as opposed to amodal symbol-processing, view of human cognition). For instance, although they are often presented as overcoming the problems illustrated by the Chinese Room argument⁶, they tend to ignore that Searle, in the original paper⁷, already rejected the “robot reply” – collecting inputs from sensors and manipulating them to produce motor outputs – as a way of achieving an AI to which genuine understanding and mental states can be attributed (see 6 for a fuller discussion).

However, this is not to say that they have no explanatory power (or indeed no utility!) at all. For example, even strongly abstracted models of sensorimotor mechanisms can provide insights into minimal requirements for the cognitive process of interest³. While not necessarily creating an account of cognitive mechanisms per se, such strategies can nonetheless constrain the search space. Similarly, theoretical models can explore how changes in embodiment might affect cognitive processes that depend on it⁸.

Overall, therefore, the realisation that human cognition is embodied to a degree that cannot be abstracted away makes “exploring the implications of ideas” with computational models much more challenging. At the same time, however, robots also create an entirely new raison d’être for cognitive models, which does not rely on explanatory insights about human cognition: application-driven models for human-machine interaction. The remainder of this paper is dedicated to the discussion of such an example.

2.Non-explanatory uses of models

Issues in explanatory power notwithstanding, technological advances in a number of areas dealing with human-machine interaction^9–11 give models of cognitive mechanisms an important reason of existence: to interact proficiently with humans, such machines need at least a rudimentary Theory of Mind (ToM); an internal model that can be used to estimate mental states of humans, in particular their intentions, expectations and predicted reactions to actions by the agent^10,12.

Additionally, to create machines with a given human ability necessitates a mechanistic model of this ability. Whether or not the model provides an adequate explanation of the mechanisms underlying the cognitive behaviour in a human is irrelevant here; it is sufficient that the postulated mechanisms can be exploited by the artificial agent. In contrast with McClelland’s take on the utility of models, such models (although of human cognitive phenomena) do not need to possess any explanatory power; appropriate behaviour according to specification when instantiated in an artificial agent, no matter how biologically implausible the underlying mechanisms, is sufficient.

A particular human phenomenon of interest here is the emergence of appropriate cognitive behaviours. It is generally not possible to fully specify the behaviour that artificial cognitive agents ought to display since that would require complete knowledge of every situation they are likely to encounter. Emergent models therefore do not specify agent behaviour axiomatically at design time; rather the agent discovers appropriate behaviour^a its...

Table of contents

• Cover
• Halftitle
• Progress in Neural Processing
• Title Page
• Copyright
• Preface
• Contents
• Patchy Connectivity and Visual Processing Asymmetries: A Neurodevelopmental Hypothesis
• Neural Binding in Letter- and Word-Recognition
• Learning Bidirectional Connections between Areas with Standard Spike-Timing-Dependent Plasticity
• A Critique of Pure Hierarchy: Uncovering Cross-Cutting Structure in a Natural Dataset
• Implementing the “Simple” Model of Reading Deficits: A Connectionist Investigation of Interactivity
• Dynamics of Word Learning in Typically Developing, Temporarily Delayed, and Persistent Late-Talking Children
• Complex Word Recognition Behaviour Emerges from the Richness of the Word Learning Environment
• How to Design Emergent Models of Cognition for Application-Driven Artificial Agents
• Competition Affects Word Learning in a Developmental Robotic System