In the project's first two years, we have started and advanced on all of the three main axes of the project: computational, experimental, and related to machine learning implementations.
Regarding the computational part of the project, we have two main scientific results. In the first study, we implemented a mean-field multiscale model of different cortical regions to investigate how their interaction produces oscillatory dynamics and traveling waves. Importantly, our model spans different scales, replicating cortical dynamics (macroscale) and laminar ones (mesoscale), allowing us to further elaborate on the predictive coding mechanisms involved in generating oscillatory traveling waves. In the same model, we also investigate the role of thalamic nuclei (i.e. the pulvinar) in coordinating cortical oscillations. In the second work, we compared different methods to quantify traveling waves in a series of connected nodes representing different brain regions. Specifically, we contrasted phase-based methods, which are widely used to quantify traveling wave propagation in neural recordings, and Granger-based methods, which measure causality between the activities in different areas. Here, through a series of simulations based on autoregressive models, we showed the difference between the two methods, demonstrating that the traveling wave direction doesn't necessarily reflect the actual flow of information (as measured by Granger-based methods).
Concerning the experimental work, we pursued different lines of research. In the first work, we investigated the relationship between traveling waves and visual attention. Remarkably, we could dissociate the waves propagating in opposite directions (either bottom-up or top-down the visual system) and relate them to different cognitive processes: sensory processing and visual inhibition. In a second study, in collaboration with a team from Lausanne (Switzerland), we showed the difference in traveling waves between schizophrenia patients and healthy controls, as predicted by the predictive coding framework. These results are particularly encouraging as they experimentally support the central hypothesis of our work, which is a relationship between traveling waves and predictive processes. A third study investigated the link between traveling waves and visual working memory. We reanalyzed publicly available data and disentangled the role of alpha-band traveling waves during working memory tasks. Lastly, ongoing work directly investigates the link between traveling waves and predictive processes. Specifically, we are analyzing physiological data (EEG and pupillometry) to identify markers of statistical learning, which is an excellent paradigm for investigating predictive processes, as it is possible to manipulate the predictability of upcoming stimuli and correlate the physiological responses with predictable (or unpredictable) events.
In the last axis, we investigate machine learning implementations based on predictive coding and oscillatory dynamics. In the first project, we developed a convolutional neural network integrating predictive coding dynamics. Similar to our simulations based on physiological models, we observed the emergence of oscillatory dynamics due to the biological constraints (like temporal delays) imposed in the architecture. In this work, we aimed to show the emergence of traveling waves in machine learning implementation to investigate whether they could be used to synchronize the activity of the network to improve their performance. Similarly, in another project, we also introduced oscillatory dynamics (without predictive coding mechanisms), taking inspiration from an influential architecture in artificial intelligence, the transformer. In this work, we introduced a global space that integrates and synchronizes the activity of the network, improving its performance over time.
