Before that, it was a computer program. That's according to a paper, which appeared in Biological Psychiatry last month, although it involved no biology, called Using Computational Patients to Evaluate Illness Mechanisms in Schizophrenia.The authors set up a neural network model, called DISCERN, and trained it to "read" stories. The nuts and bolts are, we're reassured, not something that readers of Biological Psychiatry need to worry about: "Its details, many of which are not essential in understanding this study..."
Anyway, it's basically a series of connectionist models. These are computer simulations of a large number of simple units, or nodes, which can have "activations" of varying strengths, and which have "connections" to other nodes. The model "learns" by modifying the strength of these connections according to some kind of simple learning rule.
Connectionist models are a bit like brains, in other words. A bit. They're several orders of magnitude simpler than a real brain, in several different respects. Still, they can "learn" to do some quite complicated things. You can train them to recognise faces and stuff, which is not trivial.
Anyway, DISCERN is a connectionist model of language, but it's not necessary a model of how the human brain actually learns language. Because we just have no idea how the human brain does that. We don't even know if our brain acts as a connectionist network at all, above the cellular level. Some cognitive scientists think it is, but others think that those guys are talking out of an orifice connected to their mouth, but not their mouth. Not in so many words you understand.
So they set up this system and got it to learn 28 stories, each of which consisted of multiple sentences. Some of the stories were the autobiography of a doctor - "I was a doctor. I worked in New York. I liked my job. I was good doctor" - he was not a great communicator, clearly. Others were a story about gangster ("Tony was a gangster. Tony worked in Chicago..." etc.) The network had to read these stories and then recall them.
The core of the study was that they tested to see what happened when they interfered with the program by introducing certain bugs - interfering with the activations or connections of nodes in particular parts of the model. They tried 8.
They compared the computer's performance to that of 37 actual patients with schizophrenia (or the related schizoaffective disorder) who were tested on a similar task, compared to 20 healthy controls. When the human patients came to recall the stories they'd read, they tended to make more errors of particular kinds: mixing up who did what ("agent switching"), and adding stuff that wasn't in the story ("derailment").
What they found was that DISCERN made the same kinds of errors when it was given 2 particular deficits, "working memory disconnection" and "hyperlearning". The other 6 deficits didn't cause the same pattern of findings. Hyperlearning was the best match.
They comment that
A majority of three-parameter best-fit hyperlearning simulations also recurrently confused specific agents in personal stories (including the self-representation) with specific agents in crime stories (and vice versa) in a highly nonrandom fashion.
Noteworthy was the high frequency of agent-slotting exchanges between the hospital boss, Joe, and the Mafia boss, Vito, and parallel confusions between the “I” self-reference and underling Mafia members, suggesting generalization of boss/underling relationships.
Insofar as story scripts provide templates for assigning intentions to agents, a consequence of recurrent agent-slotting confusions could be assignment of intentions and roles to autobiographical characters (possibly including the self) that borrow from impersonal stories derived from culture or the media.
Confusion between agent representations in autobiographical stories and those in culturally determined narratives could account for the bizarreness of fixed, self-referential delusions, e.g., a patient insisting that her father-in-law is Saddam Hussein or that she herself is the Virgin Mary.
It's fair to say that this is speculative. On the other hand, it's an interesting approach, and at least it's theory-based, rather than just an attempt to use ever more powerful genetic, neuroimaging and biological techniques to find differences between a patient group and a control group.
Hoffman RE, Grasemann U, Gueorguieva R, Quinlan D, Lane D, & Miikkulainen R (2011). Using computational patients to evaluate illness mechanisms in schizophrenia. Biological psychiatry, 69 (10), 997-1005 PMID: 21397213
12 comments:
Can't say I really understand their computer models, but, being a pessimist, I have to question that they can create an analogous model that reenacts a neuropsychological process that is so poorly understood in the human subject. A delusion is a symptom. People with brain injuries sometimes experience delusions. Is it the same process as psychosis? Maybe, in some cases. In other cases there are qualitative differences.
Did they also notice any "negative" symptoms, disorganization, auditory hallucinations, other signs and symptoms of schizophrenia?
I wonder how well cognitive behavioral therapy works on a dish of neurons? What about a healthy dose of thioridizine on the motherboard?
Seems to lead to more questions
You can make a model do whatever you want, the relevance is zero. If 'climate science' (never two words were less appropriate in one phrase) proved anything it is that.
Waiting with bated breath for the depressed computer. Or the computer with BPD, everytime you walk away it feels rejected.
Were do you get this stuff?
This sounds more like a simulation of dreams. The "nonsense" is quite similar.
But maybe schizophrenia is a condition in which you dream while awake? (Including bad dreams.)
This is interesting but it is still rubbish as far as understanding the human mind and its dysfunctions.
Must disagree that computer modelling is a poor method of understand brain disorders. While this may be a (relatively crude) model of delusion formation it still performed closest to that of schizophrenic patients in this particular task.
A more pressing issue is how this fits with the rest of the schizophrenia literature. If 'hyperlearning' imitates delusion formation, how does this fit with evidence of glutamate antagonists such as ketamine inducing similar delusional thoughts? Surely a glutamate antagonist would inhibit learning rather than enhance it given Hebbian models of learning.
I would have thought that without a really quite detailed level of knowledge of how the model works then mixing up "the hospital boss, Joe, and the Mafia boss, Vito" and "the “I” self-reference and underling Mafia members" is pretty difficult to interpret.
Since the model has not even been shown to be analogous to normal human speech comprehension* then the ability to reproduce errors made by patients may well tell us absolutely nothing about human disease.
* Whatever that is supposed to be when considered separate to human thought in general.
"Noteworthy was the high frequency of agent-slotting exchanges between the hospital boss, Joe, and the Mafia boss, Vito, and parallel confusions between the “I” self-reference and underling Mafia members, suggesting generalization of boss/underling relationships."
For the model to recognize these types of relationships, the authors would have had to explicitly tag these agents as possessing either these qualities the constituent elements of these qualities. In either case, it's easy to imagine post-hoc biases in the model's "memory encoder" that generate just-so results without actually reflecting the biological or theoretical underpinnings.
How these relationships are assessed by the "memory encoder" and the "story parser" has much to do with the way features are associated with lexemes. From http://nn.cs.utexas.edu/?miikkulainen:phd:
"Processing in DISCERN is based on hierarchically-organized backpropagation modules, communicating through a central lexicon of word representations. The lexicon is a double feature map, which transforms the orthographic word symbol into its semantic representation and vice versa."
I still don't buy the model thing and I refer to what Justin said above:
"it's easy to imagine post-hoc biases in the model's "memory encoder" that generate just-so results without actually reflecting the biological or theoretical underpinnings"
nicely said.
However, i do find the human subjects part of the study interesting. psychosis is essentially a thought disorder (Schizophrenia has a few added drawbacks). Delusions are one manifestation of the underlying thought problem, or so our assumptions go. The "agent switching" errors that schizophrenics tended to make is interesting, although it might be explained by general memory impairment and cognitive disorganization--that would be the simple explanation. Their leap to claiming this is evidence that delusions are a confusion of media (or other) narratives with the subject is interesting but falls short as an adequate explanation of psychotic delusions on its own. Real world schizophrenics tend to develop delusions following a handful of peculiar but predictable patterns that probably does not reflect confusion with Doctor Who.
Undergrad:
Delusions develop to preserve self-esteem, albeit in a psychotic way. How can a friggin computer have self-esteem???
pj & justin: Yes, good points. Another reason why the statement that Biological Psychiatry readers don't need to get to grips with the details, is very worrying.
Unless you understand the details - and the problem with connectionist models is that it's pretty hard to do that, because they can learn to do things in quite unexpected and unintuitive ways - you can't interpret the output.
They should study a few programming languages like how bugs work. From my understanding computational models need to be psycho specific, and not just a case of let's use a connectionist model.. did they draw that on Microsoft Word? Seems like they just wacked in whatever dude then chi-fitnessed a big mess. This should've been published in a computer magazine.. in the jokes section. The 'I love' virus according to psychiatrists can detect schizophrenia, like the Mad Hatter in Alice in Wonderland switching agents from Alice to Cheshire. The pink and purple stripes on Cheshire exacerbates a semantic confusion in the way Dr. Seuss insists his eggs are green and Sam as ham.. whatever dude.
Why is Marvin's head so big? I need to know. Aren't chips flat? I don't understand toys these days. Plastic junk connected to wires covered in fur. Quite frankly it's ethically disturbing. Then the inanimate object has the audacity to move, flap its limbs, glare at you with beady glowing eyes like it's possessed. It's disturbing. They should come up with eco-friendly toys.. like a bar of soap shaped like Pokemon, a sponge from that show Bob the sponge? malleable Gumby which can mold into a watering can, huge knitting needles doubling as a v.. kids eventually grow up! I'm curious as to who regulates the toy industry. If the children are the future, shouldn't there be careful attention to regulating the ethical aspect of toys? Not the hazardous choking warnings but the hegemony ideals toys should aspire to.. like ban superhero figurines so kids don't get depressed and delusional later in life when they end up in a 9-5 dead end job. Are there any academic journals on toys?
The analysis of error types in the people with and without schizophrenia certainly has value, but it isn't obvious how much value the computer model adds to that underlying conclusion.
Ideally, one would want to take the computer model result and go back and try to directly measure deficits in the eight different domains to see if they exhibit the same kind of sensitivity to error type (which if the theory itself used in the connnectionist computer model was evidence based, it should).
Working memory is a trait with pretty standard measurement protocols, so that wouldn't be too hard to measure directly and link to error type. Surely there is probably already even some published work on working memory in people with schizophrenia itself already.
But, what the heck is hyperlearning? Is there a test for that? Is it a bad thing? Who else has it?
Ditto, no doubt, some of the other six categories that seemed to be irrelevant.
The fact that any model or evidence would connect working memory and delusion is itself notable. What about the computer model drives that? Maybe it is too complex to understand outside a black box model. Maybe lots of neuroscience is like that.
If the future of neuroscience is ultimately a set of connections discerned via analysis of a black box rather than a true understanding of mechanisms (which is a plausible possibility for the medium term), maybe we should start funding an institute that simply spends day after day testing for connections between seemingly unrelated traits and between seemingly unrelated conditions and treatments on a trial and error basis to see what works.
After all, who would have guessed that hyperactivity can be treated with stimulants. Maybe some other totally unexpected connections exist, and finding them has value even if we have no clue how they actually work.
Post a Comment