Artificial neural networks built with feedback connections from multisensory to unisensory cortices, consistent with all-or-none models of conscious access, produced intermediate reaction times when multisensory stimuli were associated with unisensory feedback. In psychophysical testing with 29 subjects completing 10 hours of a multisensory cue-congruency task, reaction times to multisensory cues reported as unisensory fell between those of fully aware and fully unaware cues. These results suggest that graded forms of phenomenal consciousness can arise from neural networks that follow all-or-none principles.
Convergent neurons—those that receive input from multiple sources but do not integrate that information—more readily exhibit properties of consciousness than integrative neurons, contrary to the predictions of integrated information theory. In nonhuman primates under propofol anesthesia, convergent neurons showed greater neural complexity and noise correlation, and were more impacted during loss of consciousness. Neural ignition, the coactivation of primary somatosensory and ventral premotor cortex on the same trial, was more frequent in conscious states, supporting the global neuronal workspace theory. The findings directly contrast two major theories of consciousness within a single dataset.