Artificial intelligence is mastering delusional dimensions.
Having trained an artificial person to tell interesting stories, researchers from Texas tried to go further and find out the mechanism of origin of schizophrenia, opening a new chapter in the history of medicine - "virtual experimental psychiatry."
Very wrong football
Let's say you missed an important football game, but you have friends who watched the match and have already lined up to tell you the course of events. Friends are bright and wonderful people, all with their own character: one likes to embellish, the other - to lie, the third - to focus on details, the fourth - to get into the theory, the fifth confuses the chronology of goals, but remembers all the penalties, etc. it was by comparing the versions and filtering out the "noises", you will eventually be able to form a more or less realistic idea of how the events developed on the field.
To some regret, you noticed a friend N in the queue, whose versions of football matches each time more and more differ from the others, and the useful information is less and less. Once N spent an unreasonable amount of time describing the shape of the clouds over the stadium, another - he established an exact correlation between the number of goals and the number of inoperative floodlights, or even amazed at all, saying that he himself stood at the goal, the Pope was the coach of the opponents, the game was useless, and now he urgently runs away to the zoo to watch the UEFA Cup between the penguins.
For the time being, it all looked like eccentricity, until you and your friends found out that N really started a row at the zoo, then lost his job, his wife left him, and he himself leads an increasingly alienated lifestyle - he became isolated, became sad, began to collect plush penguins, does not answer calls …
In an effort to help N, but desperate to find a common language with him, you and your friends turn to specialists to find out something like the following.
Promotional video:
No matter how harmless the spectrum of schizo-like states that capture the area of "norm" is, the developed manifestations of behavioral phenomena that occur in the case of N are already related to the area of pathological dysfunctions, whose carriers can become completely incapacitated, both experiencing and causing suffering people.
When asked whether it is possible to return N from his worlds to the football club of normal people, experts will shrug. Despite many theories - from exotic (the consciousness of a schizophrenic uncontrollably spreads over a huge number of parallel stories) to more practical (genetically determined disorders in the dopamine regulation of brain neurons), the mechanisms that produce such behavior are not clear. So there is no need to count on successful correction: schizophrenic symptoms can be stopped using appropriate medications, but, firstly, not forever, and secondly, at the cost of severe side effects.
In a moment of discouragement, when it became clear that case N is a big and unsolved scientific problem, a wonderful idea comes to the bright mind of one of your friends - a specialist in artificial intelligence, neural networks and automated trading systems.
Since self-learning expert systems that mimic the work of the neural network of the brain demonstrate good results in studying the market and predicting the stock price, why not build a robot storyteller - a neural network that can remember and retell correctly, but in its own words, the history of football matches?
Further, having studied the errors and deviations that N makes when retelling matches, one can try to simulate similar deviations in the robot's nervous system by changing the parametric settings of the neural network nodes.
Gradually, using the selection method, it will be possible to establish which of the ensembles in the neural conservatory N at the mention of, say, the last victory of Spartak, instead of the prescribed club anthem, the Mendelssohn's march begins to play. Thus, by pointing out to the specialists what exactly broke in N's head and what exactly needs to be corrected there, we can increase the chances of his soonest return to the fold of adequate football fans.
DISCERN - "correct" and "schizophrenic" neurocomputer
This is the path taken by Yulai Greismann and Risto Mikkulainen from the Department of Computer Science at the University of Texas (USA), who led a mixed research group for the study of schizophrenia from specialists in artificial neural networks and employees of the Department of Psychiatry at Yale University. They built an artificial neural network called DISCERN that can memorize and retell stories, trained it, and then ran on it several hypothetical neurodysfunctions, presumably responsible for the development of schizophrenia, comparing the effects produced with the real abnormalities observed in a group of schizophrenic patients.
The results of the experiment are published in the journal Biological Psychiatry.
Unlike a classical computer that writes stories "as is" into its memory or indexes their separate common elements (say, words or even letters) in an infinitely scalable database, the DISCERN neural network perceives, remembers and reproduces information, guided by the "correct" relationships between separate elements and thus imitating the work of real brain ensembles.
Our choice of words that are more suitable to each other "in meaning" in an artificial neural network corresponds to the choice of statistically more probable connections between network elements. Fixing the "correct" relationships is achieved by "training" the neural network.
DISCERN artificial neural network architecture. Learning, memorization and reproduction of stories occurs in a chain of neuromodules. Counterclockwise from "entry": sentence analyzer, history analyzer, episodic memory, history generator, sentence generator. A separate block shows a diagram of the history generator - multilayer perceptron, inversely related to episodic memory. DISCERN neuromodules communicate with each other through a distributed representation of the "meaning" of words - fixed patterns of neural activation: only those connections are activated that have a greater weighting factor, or "more" meaning. The coefficients and patterns are fixed by training the network: in the end, receiving certain signals at the input, the network must learn to produce the correct signals at the output. The "input" signals are the first lines of the stories memorized by the network: the network must retell the rest of the story itself.
The heart of the DISCERN architecture is the story generator. Upon closer inspection, it turns out to be a classic multilayer perceptron (see illustration), familiar to us from the evolving Swiss robots. In the case of the Swiss, the task of the perceptron was to perceive external information (the “input” layer of neurons), establish the correct connections (intermediate hidden layer - “find the cube”) and produce a solution (the “output” neurons - drive up to the cube and transport).
In DISCERN, information from the episodic memory module is fed to the “input” neurons, and the “output” layer is connected simultaneously with the episodic memory and the “sentence generator”. In the course of numerous training cycles, the Swiss perceptron learned to recognize and transport cubes, in the case of DISCERN, to "correctly" reproduce the stories that the neural network remembered.
By starting training the network with random starting configurations of neural connections, you can get different types of storytellers who will tell the same story in slightly different ways - just like your friends retell the same match slightly differently.
The Texans limited themselves to thirty configurations (although there can be as many as you like), which they called “instances” (in fact, these are virtual personalities-storytellers).
Overlearning Brain Syndrome - the main cause of compulsive schizophrenic delusions?
A total of thirty copies of DISCERN learned to operate with a 159-word vocabulary, retell 28 stories ranging from three to seven simple sentences each, and distinguish “bad” stories from “good” ones. The stories were divided into autobiographical “positive” about the doctor (example: “I was a doctor / I worked in New York / I loved my job / I was a good doctor”) and “negative” about a gangster (“Tony was a gangster / Tony worked in Chicago / Tony hated his job / Tony was a bad gangster ).
At the next stage of the experiment, a group of schizophrenic patients (37 people) and a control group of healthy people (20 people) were selected. All of them were asked to listen and memorize three simple stories, and then retell them - immediately, after 45 minutes and after a week.
After analyzing the resulting texts for both groups, summary profiles were compiled, recording the observed deviations (face substitution, lexical aberrations, script changes, etc.).
Finally, at the last stage of the experiment, by changing the parametric states of individual neuroblocks in thirty virtual DISCERN storytellers, the same deviations were recorded as in real profiles.
A total of eight hypothetical neurodysfunctions were tested, presumably responsible for the development of schizophrenia.
These are disorders associated with memory (breaks in neural connections, cortical neuro noise, suppression of neuroreponse, hyperexcitation of neurons), associative dysfunctions (aberrations of semantic connections, hyperassociation, blurring and mixing of semantic signals) and signaling dysfunctions (hyper-increased brain response to a prediction error or the so-called syndrome presumably provoked by increased dopamine exposures).
As it turned out, only two scenarios provoked disorders in DISCERN narrators, similar to those observed in schizophrenics in real life.
These turned out to be memory dysfunctions and "hyperlearning syndrome", when the brain loses its ability to forget or ignore information, thus maintaining a normal proportion between noise and signal.
Effective graph of the DISCERN experiment: only two scenarios - dysfunction of memory, or Working Memory, and overlearning (top and bottom lines) - managed to adjust to the profiles of narrative deviations in the schizophrenic group (left). When adjusting to the profiles of the healthy control group, it was no longer necessary to change the parametric settings of neuromodules.
Having forgotten how to “forget”, the brain loses the ability to distinguish significant information from a huge number of exciting signals and begins to either establish connections that in reality (at least in the reality of our Universe) turn out to be inoperative (penguins playing football is a signal received when watching American cartoon "Happy Feet", not ignored by the brain and amplified), or drowns in a sea of signals, unable to organize them into a coherent story.
In the case of DISCERN storytellers, the hyperlearning syndrome (simulated by increasing the number of learning feedback loops with episodic memory) led, for example, to the fact that storytelling robots began to mix some autobiographical stories with others, substituting faces (a good doctor turned out, for example, an evil gangster) and actions (in one of the retold stories, a good doctor accused himself of committing a terrorist attack, in another he called a gangster his boss). That is, they produced illusory situations specific to the symptomatology of schizo-like delusions.
Of course, the results of virtual experiments with an artificial neural network are not yet the final proof of the correctness of one or another hypothesis explaining the development of schizophrenia. However, the very fact that an artificial neural network operating on principles similar to the brain, in some cases demonstrates behavior similar to that of real patients, opens up very exciting prospects for medicine, which has received such a powerful new tool as virtual experimental psychiatry.
All this is fine, but the paradox of the experiment lies in the fact that if the nervous system of the robots of the future imitates the human (which is exactly what is happening so far), strange subjects must also appear in their football club, who perceive an ordinary football match as an exciting excursion into a parallel reality, from which, unfortunately, there is no return yet.