Unraveling the neural correlates of dream recall: Novel insights from deep convolutional nets

People who often remember their dreams and those who rarely do show different brain activity patterns during sleep. Using deep convolutional networks on full-night EEG recordings from 36 subjects (18 high dream recall, 18 low dream recall), the authors trained separate classifiers for each sleep stage to predict which group a person belonged to. The classification accuracies revealed which sleep stages contain the best neural predictors of dream recall rate. A subject discriminator network was added as an adversary to ensure the model learned group-discriminating features rather than subject-specific ones. These features were visualized with a cue-combination Class Activation Map method. The findings are compared with prior studies, and future research directions are discussed.

Study at a glance

Abstract

Dreams and our ability to recall them are among the most puzzling questions in sleep research. Specifically, putative differences in brain network dynamics between individuals with high versus low dream recall rates, are still poorly understood. In this study, we addressed this question as a classification problem where we applied deep convolutional networks (CNN) to sleep EEG recordings to predict whether subjects belonged to the high or low dream recall group (HDR and LDR resp.). More specifically, EEG was recorded during full-night sleep from 36 subjects (18 HDR and 18 LDR). We used a deep learning framework to discover features from the EEG data that are different between the two groups (as opposed to prior knowledge-driven hand-crafted feature analysis). For each sleep stage, we trained a separate CNN to classify EEG segments into HDR and LDR group. The classification accuracies indicated which sleep stages contained the best neural predictors of dream recall rate. While training the CNN, we added a subject discriminator network and used it as adversary to restrict the CNN from learning subject-specific features and instead learn features that are subject-agnostic but allow discrimination between HDR and LDR groups. These features were subsequently visualized using a method known as cue-combination for Class Activation Map (ccCAM), inspired by [1]. These findings are compared to results obtained in previous studies [2,3] and future research directions are discussed.