Search

Perception of our sensory environment is actively constructed from sensory input and prior expectations. These expectations are created from knowledge of the world through semantic memories, spatial and temporal contexts, and learning. Multiple frameworks have been created to conceptualise this active perception, these frameworks will be further referred to as inference models. There are three elements of inference models which have prevailed in these frameworks. Firstly, the presence of internal generative models for the visual environment, secondly feedback connections which project prediction signals of the model to lower cortical processing areas to interact with sensory input, and thirdly prediction errors which are produced when the sensory input is not predicted by feedback signals. The prediction errors are thought to be fed-forward to update the generative models. These elements enable hypothesis driven testing of active perception. In vision, error signals have been found in the primary visual cortex (V1). V1 is organised retinotopically; the structure of sensory stimulus that enters through the retina is retained within V1. A semblance of that structure exists in feedback predictive signals and error signal production. The feedback predictions interact with the retinotopically specific sensory input which can result in error signal production within that region. Due to the nature of vision, we rapidly sample our visual environment using ballistic eye-movements called saccades. Therefore, input to V1 is updated about three times per second. One assumption of active perception frameworks is that predictive signals can update to new retinotopic locations of V1 with sensory input. This thesis investigates the ability of active perception to redirect predictive signals to new retinotopic locations with saccades. The aim of the thesis is to provide evidence of the relevance of generative models in a more naturalistic viewing paradigm (i.e. across saccades). An introduction into active visual perception is provided in Chapter 1. Structural connections and functional feedback to V1 are described at a global level and at the level of cortical layers. The role of feedback connections to V1 is then discussed in the light of current models, which hones in on inference models of perception. The elements of inferential models are introduced including internal generative models, predictive feedback, and error signal production. The assumption of predictive feedback relocation in V1 with saccades is highlighted alongside the effects of saccades within the early visual system, which leads to the motivation and introduction of the research chapters. A psychophysical study is presented in Chapter 2 which provides evidence for the transference of predictive signals across saccades. An internal model of spatiotemporal motion was created using an illusion of motion. The perception of illusory motion signifies the engagement of an internal model as a moving token is internally constructed from the sensory input. The model was tested by presenting in-time (predictable) and out-of-time (unpredictable) targets on the trace of perceived motion. Saccades were initiated across the illusion every three seconds to cause a relocation of predictive feedback. Predictable in-time targets were better detected than the unpredictable out-of-time targets. Importantly, the detection advantage for in-time targets was found 50 – 100 ms after saccade indicating transference of predictive signals across saccade. Evidence for the transfer of spatiotemporally predictive feedback across saccade was supported by the fMRI study presented in Chapter 3. Previous studies have demonstrated an increased activity when processing unpredicted visual stimulation in V1. This activity increase has been related to error signal production as the input was not predicted via feedback signals. In Chapter 3, the motion illusion paradigm used in Chapter 2 was redesigned to be compatible with brain activation analysis. The internal model of motion was created prior to saccade and tested at a post-saccadic retinotopic region of V1. An increased activation was found for spatiotemporally unpredictable stimuli directly after eye-movement, indicating the predictive feedback was projected to the new retinotopic region with saccade. An fMRI experiment was conducted in Chapter 4 to demonstrate that predictive feedback relocation was not limited to motion processing in the dorsal stream. This was achieved by using natural scene images which are known to incorporate ventral stream processing. Multivariate analysis was performed to determine if feedback signals pertaining to natural scenes could relocate to new retinotopic eye-movements with saccade. The predictive characteristic of feedback was also tested by changing the image content across eye-movements to determine if an error signal was produced due to the unexpected post-saccadic sensory input. Predictive feedback was found to interact with the images presented post-saccade, indicating that feedback relocated with saccade. The predictive feedback was thought to contain contextual information related to the image processed prior to saccade. These three chapters provide evidence for inference models contributing to visual perception during more naturalistic viewing conditions (i.e. across saccades). These findings are summarised in Chapter 5 in relation to inference model frameworks, transsacadic perception, and attention. The discussion focuses on the interaction of internal generative models and trans-saccadic perception in the aim of highlighting several consistencies between the two cognitive processes.