Department of Theroretical Philosophy
On the Developments in Artificial Intelligence and the Ongoing Paradigm Shift in the Cognitive Sciences
The general aim of this thesis is to investigate what developments in artificial intelligence (AI) and robotics can tell us about theoretical approaches to cognition, and vice-versa. Thereby it aims to position itself within the mutual beneficial relationship of AI and philosophy, exemplified by the slogan “understanding by building”. Two of the main research questions are: ‘to what extent are different theoretical approaches to cognition in line with exemplary implementations from robotics and AI?’ and ‘what do these exemplary implementations tell us about the different theoretical approaches to cognition?’. Three theoretical approaches to cognition are distinguished: cognitivism, computational embodied cognition and enactivism. Furthermore, four exemplary case studies from AI and robotics are picked, each exemplifying a distinctive approach to build an intelligent system. The cases to be studied are: Shakey (representational computation); Allen (subsumption architecture); evolutionary embodied agent (evolutionary robotics); and a recent success, the Atari agent (deep reinforcement learning). Though some achievements in AI and robotics from the last century had a fair amount of philosophical reflection, this is less the case for recent developments, such as deep reinforcement learning. The study of these cases resulted in some interesting insights for the theoretical approaches to cognition.
A full summary of the conclusions cannot be provided here, therefor I will highlight some of the most noteworthy results. Firstly, implementations along the lines of cognitivism and computational embodied cognition are possible, as is largely or completely exemplified in one or more of the case studies. Secondly, the case studies suggest that cognitivism is not the best way to understand all of cognition. For Shakey that is due to its limited capabilities while using representational computation; for the other implementations that is by their abilities, such as the ability to cope with a complex and changing environment, without using representational computation. A central point for enactivism is that action, cognition and perception are not only causally but also constitutively interdependent. The Atari agent exemplifies that such a dependency can be built artificially. However that does not seem to undermine the computational perspective, as the enactivists suggest. Furthermore, the Atari agent shows us that such a dependency does not depend on the body functioning as a distributor of cognitive processing or regulator of cognitive activity, thereby it exemplifies a combination of characteristics that is not found in any of the discussed approaches to cognition.
|Last modified:||08 September 2017 2.09 p.m.|