Objective. Real-time fMRI neurofeedback is a non-invasive procedure allowing the self-regulation of brain functions via enhanced self-control of fMRI based neural activation. In semantic real-time fMRI neurofeedback, an estimated relation between multivariate fMRI activation patterns and abstract mental states is exploited for a multi-dimensional feedback stimulus via real-time representational similarity analysis (rt-RSA). Here, we assessed the performances of this framework in a multi-subject multi-session study on a 3T MRI clinical scanner. Approach. Eighteen healthy volunteers underwent two semantic real-time fMRI neurofeedback sessions on two different days. In each session, participants were first requested to engage in specific mental states while local fMRI patterns of brain activity were recorded during stimulated mental imagery of concrete objects (pattern generation). The obtained neural representations were to be replicated and modulated by the participants in subsequent runs of the same session under the guidance of a rt-RSA generated visual feedback (pattern modulation). Performance indicators were derived from the rt-RSA output to assess individual abilities in replicating (and maintaining over time) a target pattern. Simulations were carried out to assess the impact of the geometric distortions implied by the low-dimensional representation of patterns’ dissimilarities in the visual feedback. Main results. Sixteen subjects successfully completed both semantic real-time fMRI neurofeedback sessions. Considering some performance indicators, a significant improvement between the first and the second runs, and within run increasing modulation performances were observed, whereas no improvements were found between sessions. Simulations confirmed that in a small percentage of cases visual feedback could be affected by metric distortions due to dimensionality reduction implicit to the rt-RSA approach. Significance. Our results proved the feasibility of the semantic real-time fMRI neurofeedback at 3T, showing that subjects can successfully modulate and maintain a target mental state, guided by rt-RSA derived feedback. Further development is needed to encourage future clinical applications.

Deep learning (DL) approaches may also inform the analysis of human brain activity. Here, a state-of-art DL tool for natural language processing, the Generative Pre-trained Transformer version 2 (GPT-2), is shown to generate meaningful neural encodings in functional MRI during narrative listening. Linguistic features of word unpredictability (surprisal) and contextual importance (saliency) were derived from the GPT-2 applied to the text of a 12-min narrative. Segments of variable duration (from 15 to 90 s) defined the context for the next word, resulting in different sets of neural predictors for functional MRI signals recorded in 27 healthy listeners of the narrative. GPT-2 surprisal, estimating word prediction errors from the artificial network, significantly explained the neural data in superior and middle temporal gyri (bilaterally), in anterior and posterior cingulate cortices, and in the left prefrontal cortex. GPT-2 saliency, weighing the importance of context words, significantly explained the neural data for longer segments in left superior and middle temporal gyri. These results add novel support to the use of DL tools in the search for neural encodings in functional MRI. A DL language model like the GPT-2 may feature useful data about neural processes subserving language comprehension in humans, including next-word context-related prediction.