Neural networks programmed with an inability to “forget” information show schizophrenia-like symptoms, according to a new experiment by Ralph Hoffman of the Yale University School of Medicine, Uli Grasemann at the University of Texas at Austin, and colleagues.

The researchers taught a series of simple stories to a neural network programmed to learn and answer questions about narratives, though with an imperfect memory. When they decreased the program’s ability to forget, it started giving answers resembling those given by humans with schizophrenia.

The answers contained dissociated sentences, digressions, and delusions—at one point the computer claimed responsibility for a terrorist bombing. Answers also typically included incoherent jumbles of elements from the various stories it had been taught.

The “hyperlearning” hypothesis

According to the researchers, the finding lends support to the “hyperlearning” hypothesis — that schizophrenia may in part be caused by an inability to forget information, depriving the brain of its ability to attach meaning to new information. Excessive releases of the neurotransmitter dopamine — which the researchers modeled by decreasing the neural network’s ability to forget — are thought to be involved in the process.

Ref.: Ralph E. Hoffman, Uli Grasemann, Ralitza Gueorguieva, Donald Quinlan, Douglas Lane, Risto Miikkulainen, Using Computational Patients to Evaluate Illness Mechanisms in Schizophrenia, Biological Psychiatry, May 15, 2011 (Vol. 69, Issue 10, Pages 997-1005)