goals of paper
DESCRIPTION
Goals of paper. Create a neural network which simulates story comprehension Determine what parts of the network are damaged to produce schizophrenic behavior. Steps to understanding a story. Identify each word (lexical access) Determine role in sentence of each word Who does what to whom? - PowerPoint PPT PresentationTRANSCRIPT
Goals of paper
• Create a neural network which simulates story comprehension
• Determine what parts of the network are damaged to produce schizophrenic behavior
Steps to understanding a story
• Identify each word (lexical access)• Determine role in sentence of each word– Who does what to whom?
• Relate sentence to the rest of the story– Use scripts and schemas to fill in gaps and make
inferences– Summarize key points of story
Two example “stories”I was a doctorI worked in New-YorkI liked my jobI was good doctor
Tony was a gangsterTony worked in ChicagoTony hated his jobTony was a bad gangster
Model for story comprehension
Comparing performance of model to unimpaired humans
Symptoms of schizophrenia
• Disorganized thought processes• Attributing acts to others or oneself
incorrectly• Dysfunctional executive disorder
Agent slotting error: Claiming incorrectly that an agent had a role in an event.eg1. The girl gave the old man the flowers is wrong.
correct: The old man gave the old man the flowers. eg2. The cop arrested me for speeding. correct: The cop arrested Vince for speeding.
Lexical misfire: incorrect words used with different meaning from story. eg. “wispy old man” “whispering man”
Derailment: entire clause of meaning is different from the story.Eg. A girl was sitting on the bus and he noticed her looking at his eyes.
Conclusions
• Computational models can be used to specify what parts of brain network break down during disorders
• Hyperlearning predicted schizophrenic behavior the best.– Exaggerated backpropagation prediction error
signaling leads to over correction, and reduces the separation between stories.
Limitations
• Only part of story memory process simulated• The network’s memory is too good! (over 95%
accuracy)• Cannot simulate unimpaired performance• Only simulates some schizophrenic behavior• Will it scale up to encode more information?