Layer-specific cortical feedback dynamics - human ultra-high resolution functional brain imaging for predictive brain functions

Driving to work, you instinctively process your surroundings whilst at the same time imagining your morning meeting. My ambition is to understand how the visual parts of the brain contribute to this feature of intelligence: associating perceived experience with internal mental models of the world, and making future predictions. This cognitive capacity of the human brain has been suggested as one of the mental abilities that in its temporal extent and flexibility separates us from animals. I propose human brains use distinct information processing streams for perceiving the present moment and for mentally simulating consequences of future actions for goal-directed behaviour over longer durations. As such, we have evolved neuronal computations that contextualise precise spatiotemporal sensory inputs, but also mentally decouple from online perception to evaluate sensory states over longer time frames. The 'Predictive Brain' theoretical framework offers a scheme to understand these processes. The framework's central tenet is that the brain performs inference and prediction to process its highly structured and dynamic environment, inferring regularities over different temporal scales to form abstract predictions evaluating future surroundings unfolding over time.

The 'Predictive Processing' framework has transformed neuroscience in the 21st century by offering an overarching theoretical framework guiding empirical brain research. The framework is experimentally tractable, and we can test it with advanced neuroscientific methods to assimilate the framework's hypothetical principles with data spanning multiple levels of brain organisation and function. Predictive processing accounts describe how brains learn in their environments by training neurons to generate internal models that explain the world. Since neuronal networks indirectly access the world via sensory pathways, internal models are optimised by forming predictions of sensory events and comparing them with actual sensory signals. The residual, or surprising, information (prediction error) is computed and processed upwards in hierarchical cortical levels where it is used to revise mental models. I have developed stimulation paradigms and fMRI approaches towards establishing how predictions are organised in human cortical microcircuits. Such brain imaging data are proving essential to constrain computational models, biologically inspired artificial intelligence and invasive neuronal recordings in primates and rodents.

My proposal outlines a novel hypothesis to be tested in mesoscale brain imaging. When sensory information is processed through cortical areas, higher areas successively feed back sensory predictions to lower areas derived from our prior expectations. Hence predictive brain signals must anticipate the rich spatial and temporal structure of sensory processing. For example, the early visual system receives a constant flow of sensory signals, and the timing of predictive feedback processing must therefore be compatible with the temporal dynamics of feedforward neural activity in order for the brain to update predictions of our perceptions over time. However, brains not only need to act in their current environment but need to plan future behaviours. I propose that cortical predictive feedback has a behaviourally relevant temporal structure, for representing the present moment or future representations. Using pioneering, ultra high-field, high-resolution human functional brain imaging, I will investigate temporal codes of predictive feedback in early visual cortical microcircuits. I will apply paradigms testing temporal predictions for perception (i.e. involving feedback in the millisecond range, such as motion illusions), for cognition (i.e. requiring feedback over seconds such as planning a route navigation), and for mental imagery (i.e. envisaging an object not currently present).

Intelligent behaviour has real-time constraints such as recognising objects, but also "offline" cognitive determinants such as mentally projecting oneself in time. We are highly dependent on our visual systems for these functions. I propose a novel conceptual model of early visual cortex (EVC) multiplexing perceptual information with predictive information about the world beyond our current view. Cognitive influences on early vision implement functions including working memory and imagery, implicating EVC as a cognitive blackboard. These accounts need to be extended by testing broader feedback coding schemes, and if early visual neurons represent not only current but future behavioural contexts. My proposal drives a new conceptualisation of EVC, contributing to prospection, a cognitive process offered as a defining facet of human intelligence.
Central to my proposal is that visual cortex forms a hierarchical architecture coding for increasing high level statistics of temporal structure in visual inputs. Temporal features are integrated with expanding complexity upwards in the cortical hierarchy, and these areas feed back internal models with temporal codes. I will use high field, high resolution brain imaging (including line scanning fMRI with a spatial resolution of 200 microns and temporal resolution of 100ms) and computational modeling to test if human visual microcircuits process temporal feedback information at sub-millimetre spatial scales on a sub-second scale intrinsic to perception, but also in time frames for cognition. We will measure topographic signatures of higher areas that have larger temporal expansion and use different coding spaces in EVC; (1) motion network to test how hMT/V5+ feedback to V1 communicates predictive models over a few hundred milliseconds, (2) place coding in the hippocampus feeds back to EVC scene information of upcoming rooms over seconds, and (3) imagined objects in IT are fed back to EVC independently of imminent appearance.