The world around us is full of sensory experiences, which follow repetitive rules. Through exposure to reoccurring stimuli, we are able to learn patterns and form predictions about future events. Predictive traces manifest in multiple stages of processing in the brain, with the most prevalent theory being that brain regions that are ‘high’ in a cortical hierarchy, (frontal areas), propagate predictions to ‘lower’ regions, (sensory areas). One way to study predictive processes is through omission paradigms: in a sequence of stimuli -sounds- a regularly repeated stimulus is rarely omitted, eliciting a neural response which is similar to the response to the actual stimulus.
Omission responses have been reported at multiple levels of a neural hierarchy. Sensory cortices show similar patterns of neural responses to omitted as to experienced stimuli. Additionally, omission responses have been reported in prefrontal and subcortical regions. This raises the question of whether omission responses originate from one global mechanism or whether independent predictive traces coexist at various levels of processing. While omissions provide a window of investigation on the neural mechanism underlying predictions, it remains unknown how omission responses change while learning a new regularity.
We hypothesize that omission responses are generated by sensory areas during learning, and by higher cognitive areas (i.e. prefrontal) when a model of the environment has been built (after learning).
The proposed project capitalizes on state-of-the-art magnetoencephalography (MEG) and machine learning techniques to achieve two goals: investigate the neural mechanisms that underlie omission responses across cortical hierarchies and study how these are modulated by learning.
MEG measures magnetic activity of the brain in a fine temporal resolution. Unlike other techniques like electroencephalography, MEG is sensitive to localizing the sources of neural activity. Unfortunately, MEG is a technique that does not exist in Switzerland, which led to the proposed collaboration with the MEG center in Russia.
We have designed an experiment which consists of presenting participants with sound sequences at different predictability levels, and including rare omissions. In March 2020 we collected preliminary MEG data in N=5 participants, which showed that: (a) the identity of omitted sounds can be decoded using machine learning models trained to discriminate responses to sounds; (b) decoding is modulated by learning and is highest in a fully learned sequence.
These first results strongly encourage the continuation of data collection over a larger cohort of N=25 participants, following standards in the MEG field. Acquiring this dataset will allow to: (a) confirm our preliminary findings on a larger cohort (b) identify brain sources (c) delineate the predictive network in order to formulate tailored protocols to test the network causality, by employing non-invasive magnetic stimulation to modulate the formation of sensory predictions. These follow-up studies will be the subject of future funding applications with all applicants.
In terms of clinical application, the completion of the proposed project will improve our understanding of pathological conditions whose neural predictive mechanisms are impaired, such as patients with disorders of consciousness and a host of neuropsychiatric disorders such as schizophrenia.
University of Bern:
University of Zurich:
Higher School of Economics (HSE) in Moscow: