Search

Computationally modelling human level cognitive abilities is one of the principal goals of artificial intelligence research, one that draws together work from the human neurosciences, psychology, cognitive science, computer science, and mathematics. In the past 30 years, work towards this goal has been substantially accelerated by the development of neural network approaches, at least in part due to advances in algorithms that can train these networks efficiently [Rumelhart et al., 1986b] and computer hardware that is optimised for matrix computations [Krizhevsky et al., 2012]. Parallel to this body of work, research in social robotics has developed to the extent that embodied and socially intelligent artificial agents are becoming parts of our everyday lives. Where robots were traditionally placed as tools to be used to improve the efficiency of a number of industrial tasks, now they are increasingly expected to emulate humans in complex, dynamic, and unpredictable social environments. In such cases, endowing these robotic platforms with (approaching) human-like cognitive capabilities will significantly improve the efficacy of these systems, and likely see their uptake quicken as they come to be seen as safe, effective, and flexible partners in socially oriented situations such as physical healthcare, education, mental well-being, and commerce. Taken together, it would seem that neural network approaches are well placed to allow us to bestow these agents with the kinds of cognitive abilities that they require to meet this goal. However, the nascent nature of the interaction of these two fields and the risk that comes along with integrating social robots too quickly into high risk social areas, means that there is significant work still to be done before we can convince ourselves that neural networks are the right approach to this problem. In this thesis I contribute theoretical and empirical work that lends weight to the argument that neural network approaches are well suited to modelling human cognition for use in social robots. In Chapter 1 I provide a general introduction to human cognition and neural networks and motivate the use of these approaches to problems in social robotics and human-robot interaction. This chapter is written in such a way that readers with no technical background can get a good understanding of the concepts that are at the center of the thesis' aims. In Chapter 2, I provide a more in-depth and technical overview of the mathematical concepts that are at the heart of modern neural networks, specifically detailing the logic behind the deep learning approaches that are used in the empirical chapters of the thesis. While a full understanding of this chapter requires a stronger mathematical background than the previous chapter, the concepts are explained in such a way that a non-technical reader should come out of it with a solid high level understanding of these ideas. Chapters Chapter 3 through Chapter 5 contain the empirical work that was carried out in order to attempt to answer the above questions. Specifically, Chapter 3 explores the viability of using deep learning as an approach to modelling human social-cognitive abilities by looking at the problems of subjective psychological stress and self-disclosure. I test a number of "off-the-shelf" deep learning architectures on a novel dataset and find that in all cases these models are able to score significantly above average on the task of classifying audio segments in relation to how much the person performing the contained utterance believed themselves to be stressed and performing an act of self-disclosure. In Chapter 4, I develop the work on subjective-self disclosure modelling in human-robot social interaction by collecting a much larger multi modal dataset that contains video recorded interactions between participants and a Pepper robot. I provide a novel multi-modal deep learning attention architecture, and a custom loss function, and compare the performance of our model to a number of non-neural network approach baselines. I find that all versions of our model significantly outperform the baseline approaches, and that our novel loss improves on performance when compared to other standard loss functions for regression and classification problems for subjective self-disclosure modelling. In Chapter 5, I move away from deep learning and consider how neural network models based more concretely on contemporary computational neuroscience might be used to bestow artificial agents with human like cognitive abilities. Here, I detail a novel biological neural network algorithm that is able to solve cognitive planning problems by producing short path solutions on graphs. I show how a number of such planning problems can be framed as graph traversal problem and show how our algorithm is able to form solutions to these problems in a number of experimental settings. Finally, in Chapter 6 I provide a final overview of this empirical work and explain its impact both within and without academia before outlining a number of limitations of the approaches that were used and discuss some potentially fruitful avenues for future research in these areas.

Neoliberalism is by now understood as both an approach to government and also the defining political movement of the last fifty years. In both instances, neoliberalism is built on the assumption that the state is not ideally placed to create economic growth or provide a social safety net, and that instead private companies, private individuals, and, most importantly, unhindered markets are best placed to generate economic growth and provide optimum conditions for social equality. Rather than delivering on these promises, however, neoliberalism has fused individual self-interest with the most devastating effects of capitalism, thereby increasing inequalities, creating newly excluded populations, generating widespread precarity and delivering mass unemployment. This thesis utilises the practices and traditions of clown, as a means of critiquing neoliberal hegemony. In so doing it brings a traditional popular performance mode into conversation with politics. In particular, by examining neoliberal logic from the point of view of the clown, in practice and in theory, its intention is to rescue failure from its current condition as always and everywhere to be avoided. It argues that since neoliberalism celebrates only success, failure has been robbed of its productive potential in both social and political terms. The clown is ideally positioned to redeem failure, because of his expert skill in this area. Finally, this thesis maps the relationships between different discourses - politics and popular performance - with the aim of generating insights that have not yet been articulated.