Deep in predictions: Unravelling visual perception with deep neural networks inspired by hierarchical predictive coding

The thesis presents PrediRep, a new type of deep learning network that imitates how the brain processes visual information, based on the theory of hierarchical predictive coding (hPC). hPC suggests that the brain learns by predicting sensory inputs using two kinds of neurons: representation neurons that capture sensory data, and error neurons that detect differences between predictions and actual inputs. PrediRep uses deep learning to create a multi-layered network that processes video data, which closely resembles real-world stimuli. This network structure mirrors the brain's layered organization and integrates both representation and error units at each level, following hPC principles more closely than other networks. Although PrediRep isn’t focused on predicting video frames, its performance matches other models while using fewer parameters. PrediRep consistently displays several neural phenomena observed in the visual cortex, such as endstopping and orientation selectivity, reinforcing its alignment with hPC. It provides researchers with a tool to explore visual processing and the mechanisms of hPC.