Abstract

This dissertation focuses on generative models in the Predictive Processing framework. It is commonly accepted that generative models are structural representations; i.e. physical particulars representing via structural similarity. Here, I argue this widespread account is wrong: when closely scrutinized, generative models appear to be non-representational control structures realizing an agent’s sensorimotor skills. The dissertation opens (Ch.1) introducing the Predictive Processing account of perception and action, and presenting some of its connectionist implementations, thereby clarifying the role generative models play in Predictive Processing. Subsequently, I introduce the conceptual framework guiding the research (ch.2). I briefly elucidate the metaphysics of representations, emphasizing the specific functional role played by representational vehicles within the systems of which they are part. I close the first half of the dissertation (Ch.3) introducing the claim that generative models are structural representations, and defending it from intuitive but inconclusive objections. I then move to the second half of the dissertation, switching from exposition to criticism. First (Ch.4), I claim that the argument allegedly establishing that generative models are structural representations is flawed beyond repair, for it fails to establish generative models are structurally similar to their targets. I then consider alternative ways to establish that structural similarity, showing they all either fail or violate some other condition individuating structural representations. I further argue (Ch.5) that the claim that generative models are structural representations would not be warranted even if the desired structural similarity were established. For, even if generative models were to satisfy the relevant definition of structural representation, it would still be wrong to consider them as representations. This is because, as currently defined,structural representations fail to play the relevant functional role of representations, and thus cannot be rightfully identified as representations in the first place. This conclusion prompts a direct examination of generative models, to determine their nature (Ch.6). I thus analyze the simplest generative model I know of: a neural network functioning as a robotic “brain” and allowing different robotic creatures to swiftly and intelligently interact with their environments. I clarify how these networks allow the robots to acquire and exert the relevant sensorimotor abilities needed to solve the various cognitive tasks the robots are faced with, and then argue that neither the entire architecture nor any of its parts can possibly qualify as representational vehicles. In this way, the structures implementing generative models are revealed to be non-representational structures that instantiate an agent’s relevant sensorimotor skills. I show that my conclusion generalizes beyond the simple example I considered, arguing that adding computational ingredients to the architecture, or considering altogether different implementations of generative models, will in no way force a revision of my verdict. I further consider and allay a number of theoretical worries that it might generate, and then briefly conclude the dissertation.