Thesis defense Danaja Rutar (Donders series 589)
28 February 2023
Promotor: prof. dr. S. Hunnius
Co-promotor: dr. J. Kwisthout
I model the world, therefore I am. Representing and changing the structure of the world in generative models
My thesis explored two research topics within the theoretical framework of predictive processing. I examined how information is represented in generative models and how generative models change in structure as a consequence of learning. One of the central tenets of predictive processing is that generative models make predictions based on their internal representation of the world. What exactly do generative models represent though? In Chapter 2 I investigated how the format od represented information changes depending on how the sensory input is structured. In Chapter 3 I further argued that two functional features of representations, coding for structural similarity and operating in a decoupled manner, are gradual and that modulating the degree of decoupling and structural similarity is conducive to behavioural success of learners. In predictive processing, learning has been mostly formalised as Bayesian model updating. Whilst the latter provides important insights into how humans change the degree of belief in existing hypotheses in light of new evidence, it cannot explain how new hypotheses are incorporated in generative models (i.e., structure learning). In Chapter 4 I argued that the only available computational implementation of structure learning in predictive processing fails to capture what is most distinctive about structure learning – the change in structure of generative models. We put forward an improved version of structure learning theory, by proposing a set of operations for changing the structure of generative models. Chapter 5 investigated whether Bayesian model updating and structure learning are not only computationally distinct but can also be empirically distinguished based on the prediction error dynamics. The results show preliminary evidence that structure learning and Bayesian model updating are indeed empirically differentiable. Finally, we discussed what implications the findings presented in the thesis have for understanding adult and child higher cognition.