Researchers from the University of Texas at Austin, USA, have made a learning computer algorithm that show signs of schizophrenia.
The computer program, called DISCERN (Distributed Script Processing and Episodic Memory Network), was published within Biological Psychiatry in April 2011 by professor Risto Miikkulainen, and his graduate student Uli Grasemann.
“One of the stories we used is, ‘I wanted to drive to the airport.’ The sentence was fed to DISCERN as a sequence of individual words, i.e. first ‘I,’ then ‘wanted’ etc. Each word is represented as a pattern of numbers; similar words (in terms of word meaning) have similar patterns. The patterns were fed to the first neural network (the "sentence parser"), whose output was then fed to another network, and so on,” said Grasemann.
Back in the ‘real world’ researchers compared 20 healthy control subjects and 37 patients with schizophrenia. They analyzed the mechanisms that best reproduced narrative breakdowns between healthy subjects and schizophrenic patients.
This particular theory is called ‘hyperlearning’ and states that the biochemical dopamine is crucial in encoding the importance, or salience, of experience in humans. Altering levels of dopamine can therefore have a strong impact on how we interpret the world. The researchers then tested whether they could inflict schizophrenic-like behavior or ‘delusion-like narratives’ in DISCERN’s answers.
The researchers simulated an extreme release of dopamine in DISCERN by increasing the system's learning rate. The system was told not to be able to forget. Being able to ignore information is crucial in human learning. "What we found is that if you crank up the learning rate in DISCERN high enough, it produces language abnormalities that suggest schizophrenia,” said Grasemann.
DISCERN’s increased learning rate caused its output to be confused, putting itself in fantastical stories that had no bearing on the coherent stories it was being fed. In one instance it claimed to be responsible for a terrorist bombing.
Ralph Hoffman, professor of psychiatry at Yale School of Medicinesaid, “simulated hyperlearning provided a new explanation for how these fixed delusions can occur, namely where hyperlearning produces intermingling of character references that come to be stored as memories. The best approach would be to use neuroimaging methods to characterize the rate of consolidation of narrative memories into long-term storage.”
DISCERN could also help study more than just schizophrenia. “We could develop plausible models of bipolar disorder by developing a separate emotion processing module,” said Hoffman.