Frontiers in Human Neuroscience
January 3, 2024
Keisuke Suzuki, Anil K. Seth, David J. Schwartzman
20 citations
Visual hallucinations differ substantially depending on their cause, such as neurodegenerative disease, visual loss, or psychedelic drugs. Using a deep neural network approach called computational (neuro)phenomenology, researchers identified three key dimensions that distinguish these hallucinations: realism (how true-to-life they seem), spontaneity (how much they depend on sensory input), and complexity. By tuning the network along these dimensions, they generated synthetic hallucinations characteristic of each cause. Two studies with patients having Parkinson's disease, Lewy body dementia, or Charles Bonnet syndrome, and people with recent psychedelic experience, confirmed that these synthetic images matched the phenomenology reported by each group. The findings show that a neural network model can capture the distinctive visual features of hallucinations from different origins.
bioRxiv Preprint Server
November 3, 2017
Keisuke Suzuki, Warrick Roseboom, David J. Schwartzman et al.
2 citations
preprint
A tool called the Hallucination Machine simulates visual hallucinatory experiences using deep convolutional neural networks and panoramic virtual reality videos of natural scenes. It induces visual phenomenology qualitatively similar to classical psychedelics, but does not evoke the temporal distortion commonly associated with altered states. This technique allows researchers to study altered consciousness without the confounding physiological and cognitive effects of psychoactive substances or psychopathological conditions, offering a valuable method for consciousness science and psychiatry.
Frontiers in Psychology
May 20, 2026
Keisuke Suzuki
Altered states of consciousness—such as hallucinations, psychedelic experiences, and ego dissolution—differ qualitatively, but no unified computational framework describes what varies and along which dimensions. This paper proposes the C × G × D framework, drawing on three functional roles in deep neural networks: a Classifier (C) that extracts features from sensory input, a Generator (G) that synthesises internal representations, and a Discriminator (D) that judges whether a representation originates externally or internally. Phenomenological differences across altered states are redescribed as variations in objective functions, constraints, and thresholds of these components.
Frontiers in psychology
January 1, 2026
Felix Woolford, Keisuke Suzuki
Intentional binding, a measure of how people perceive time between their actions and outcomes, was tested in individual, human-computer, and human-human joint button-pressing tasks using haptic devices. Contrary to expectations from we-agency theory, the overall strength of binding did not differ between partner types. However, within human pairs, participants who reported a stronger sense of agency showed stronger binding, linked to leader-follower movement dynamics. No such link appeared in human-computer interactions. The findings indicate that temporal binding primarily reflects sensorimotor predictability rather than social context or intentionality, and may serve as a signature of how partners co-regulate their actions.