Dreaming is generally thought to be generated by spontaneous brain activity<br />
during sleep with patterns common to waking experience. This view is supported<br />
by a recent study demonstrating that dreamed objects can be predicted from<br />
brain activity during sleep using statistical decoders trained with<br />
stimulus-induced brain activity. However, it remains unclear whether and how<br />
visual image features associated with dreamed objects are represented in the<br />
brain. In this study, we used a deep neural network (DNN) model for object<br />
recognition as a proxy for hierarchical visual feature representation, and DNN<br />
features for dreamed objects were analyzed with brain decoding of fMRI data<br />
collected during dreaming. The decoders were first trained with<br />
stimulus-induced brain activity labeled with the feature values of the stimulus<br />
image from multiple DNN layers. The decoders were then used to decode DNN<br />
features from the dream fMRI data, and the decoded features were compared with<br />
the averaged features of each object category calculated from a large-scale<br />
image database. We found that the feature values decoded from the dream fMRI<br />
data positively correlated with those associated with dreamed object categories<br />
at mid- to high-level DNN layers. Using the decoded features, the dreamed<br />
object category could be correctly identified by matching them to the averaged<br />
features for multiple candidate categories. The results suggest that dreaming<br />
recruits hierarchical visual feature representations associated with objects,<br />
which may support phenomenal aspects of dream experience.