The Effect of Structural (mis)alignment of Temporal Generative Models on Prediction Error

Keywords
Loading...
Thumbnail Image
Issue Date
2021-06-01
Language
en
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
It is often assumed that predictive processing agents can e ectively learn the dynam- ics of an environmental variable (e.g. temperature) that changes over time at multiple timescales using an internal hierarchical generative model, where higher layers represent slower, more abstract, processes than lower layers. In this research paper, we investigate how accurately a perceptive inference agent is able to learn and predict these dynamic environment variables using such a model. More speci cally, we answer the follow- ing question: What is the e ect of structural (mis)alignment of hierarchical temporal generative models on prediction error? To do this, we used three scenarios. In the rst scenario, a generative model was structurally aligned with the generative process. In the second scenario, we created a generative model that had one or more layers removed or added with respect to the structurally aligned model. In the third scenario, the generative model had a layer which was smaller or larger in comparison to the structurally aligned agent. In particular, we looked at how well the agents managed to learn these dynamics in terms of prediction error minimization. We found that generally models with less layers learn a more accurate prediction of the environment, but only if the number of unique state combinations remains the same as with the structurally aligned agent. If a layer has less states than the structurally aligned equivalent, this layer and the layers above it are not able to learn a represen- tation of the process at all. If a layer has more states, however, this layer can learn a representation but only if the layer size is an integer multiple of the structurally aligned equivalent.
Description
Citation
Faculty
Faculteit der Sociale Wetenschappen