Neural representation of consciously seen and unseen information.

Machine learning can decode whether someone consciously perceives a near-threshold visual stimulus from their EEG brain signals, and using time-frequency representations of the signal improves decoding accuracy over raw voltage, especially in theta and alpha frequency bands. In a task where participants viewed faint Gabor gratings and reported whether they saw them, the model successfully detected the presence or absence of a stimulus and the participant's subjective perception, but could not decode the grating's orientation. Unconscious processing of unseen stimuli was evident both behaviorally and neurally, but these unconscious representations were less stable over time and appeared only at early perceptual stages around 100 milliseconds and during response preparation.

Study at a glance

Abstract

Machine learning (ML) techniques have steadily gained popularity in Neuroscience research, particularly when applied to the analysis of neuroimaging data. One of the most discussed topics in this field, the neural correlates of conscious (and unconscious) information, has also benefited from these approaches. Nevertheless, further research is still necessary to better understand the minimal neural mechanisms that are necessary and sufficient for experiencing any conscious percept, and which mechanisms are comparable and discernible between conscious and unconscious events. The aim of this study was two-fold. First, to explore whether it was possible to decode task-relevant features from electroencephalography (EEG) signals, particularly those related to perceptual awareness. Secondly, to test whether this decoding could be improved by using time-frequency representations instead of voltage. We employed a perceptual task in which participants were presented with near-threshold Gabor stimuli. They were asked to discriminate the orientation of the grating, and report whether they had perceived it or not. Participants' EEG signal was recorded while performing the task and was then analysed by using ML algorithms to decode distinctive task-related parameters. Results demonstrated the feasibility of decoding the presence/absence of the stimuli from EEG data, as well as participants' subjective perception, although the model failed to extract relevant information related to the orientation of the Gabor. Unconscious processing of unseen stimulation was observed both behaviourally and at the neural level. Moreover, contrary to conscious processing, unconscious representations were less stable across time, and only observed at early perceptual stages (~ 100 ms) and during response preparation. Furthermore, we conducted a comparative analysis of the performance of the classifier when employing either raw voltage signals or time-frequency representations, finding a substantial improvement when the latter was used to train the model, particularly in the theta and alpha bands. These findings underscore the significant potential of ML algorithms in decoding perceptual awareness from EEG data in consciousness research tasks.