Schizophrenic Computer Points to New Theory of Disease
Software pays undue attention. Does a diseased mind do the same?
Illustration: Randi Silberman Klett/iStockphoto
Researchers at the University of Texas at Austin have modeled the disease in a program called DISCERN in an attempt to prove a long-standing theory of how schizophrenia works. But what they came away with was a completely new clinical hypothesis.
DISCERN, short for distributed script processing and episodic memory network, was built to process and recall simple narratives. With the original settings, it can digest a narrative, retain it, and reproduce the story in its "own" words. Every script fed to DISCERN runs through multiple modules that simulate networks in the brain. As the information flows through a semantic memory module, a sentence parser, and a story parser, the system learns what words mean by looking statistically at how they operate in sentences. In a simulation of working memory, it identifies different aspects of the story to change the likelihood that it will retain a specific detail, then rebuilds the parts it does remember back into a story. In this way, DISCERN models the brain both as a semantic and emotional processor.
Doctors listen very closely to language in their schizophrenic patients and to how they recall stories. People with the disease often confuse who did what to whom, resulting in wild tales that become autobiographical, in which they place themselves at the center of fantastic events. Another type of schizophrenia causes patients to combine completely different stories into a single narrative.
"We’re so tuned to language," says Uli Grasemann, a lead author on the study. "We use it for diagnosis. We may as well use it to search for compelling models."
Teaming up with Ralph Hoffman, a psychiatrist from Yale University, Grasemann began tampering with DISCERN and diagnosing it as though it were a human being. Hoffman initially wanted to bolster a theory that links schizophrenic behavior to a breakdown of communication between areas of the brain controlling working memory. To do this, the team reduced the connectivity in DISCERN’s story generator. The results were very disappointing.
The researchers tried to damage their simulated brain in other ways. In all, they simulated seven hypothetical causes of the disease, but none of the injuries produced the kind of language they were looking for.
Grasemann then decided to train DISCERN’s story generator for a longer period before starting the next experiment, hoping this would bring some behaviors out of the shadows. He also weighted the extra training more heavily so that everything DISCERN learned during this interval would take on more importance. What he produced was a schizophrenic computer.
After digesting a narrative about a terrorist named Tony who drove to city hall and detonated a bomb, DISCERN retold the story in the first person. DISCERN was just as eager to adopt the life of a crime boss named Vito or to thrust its assigned identity onto another character.
But when Grasemann enhanced the training in the system’s memory encoder, he noticed very different symptoms. Suddenly, DISCERN began losing the thread of stories, hopping erratically between narratives. By changing the same parameter in different places, Grasemann had modeled both the delusions and the disordered speech typical of schizophrenic patients.
The researchers were quickly convinced that they’d found something important. "In general, when you develop models, if you have enough parameters, you can make them do just about anything. If it’s unplanned and unanticipated, it makes it much more powerful," says Hoffman.
While setting out to prove one theory, Hoffman had stumbled upon a completely new one that he calls the "hyperlearning hypothesis." The theory proposes that a period of intense learning, during which the brain assigns an unwarranted importance to new information, can set off delusions and scattered language.
This is very similar to a well-known theory that increased dopamine transmission contributes to schizophrenic symptoms. In some parts of the brain, dopamine seems to have a role in marking relevant information, explains Alexander Arguello, a neuroscientist at Princeton University. When this system becomes hyperactive, an individual may begin to feel that everything is deeply but mysteriously relevant. In this theory, delusions may be a strategic attempt to rationalize this state of heightened awareness.
"It’s not just that you’re remembering things better. You’re remembering things that aren’t related to each other. You’re making the wrong associations between things," says Arguello.
Although other computer systems have been used to model schizophrenic symptoms, DISCERN is the first to look at changes in language and storytelling. "This study is the only one that looks at how abnormal salience might lead to some of the language deficits observed in patients," says Arguello, "providing further support that prediction-error learning deficits might be involved in schizophrenia pathophysiology."
But for now, Hoffman and Grasemann are using DISCERN as a brainstorming tool, and both acknowledge that proof of the hyperlearning hypothesis can come only from the lab or the clinic.
"It has to do with the strength of the claim," Grasemann says of their path to a proven theory. "At this point, we only have to be interesting."