A strange idea crossed my mind a year and a half ago when I worked on an article about conscious artificial intelligence. I knew so little about the subject then that I just trusted my gut feeling. It insisted that there must be a link between human consciousness and narratives. In particular, I got interested in fairy tales which origins were traced by scientists at least to the Bronze Age. Stories, some research claimed, were also the way of encoding of human values in human brains using the Default Mode Network of the brain for that purpose.
“Maybe, just maybe,” I thought, “we can use stories, most ancient fairy tales in particular, to make modern humans more human friendly and to also develop human consciousness in machines to make them friendly to humans as well.” If I would know then at least 5% of what I know now, I would never ever even dare to think about such crazy, stupid, bizzare ideas. If I would stop then, I would regret about it all the rest of my life.
Surprisingly, some of my friends who were not stupid, neither crazy, suddenly, became fascinated with this idea. Together we just kept going on without a slightest idea of where this path will take us. We soon identified some research, in which stories were used to teach human values to artificial agents. It looked like we were on the right track.
Last July a worldwide famous scientist in the area of deep learning got interested in our idea of teaching ethical values to humans and machines by using storytelling. It was a miracle and as any miracle it took us by surprise. We had no idea of how we can produce a corpus of high quality training data that was needed to train a deep neural network the basics of human friendly behavior. Written texts of 50 fairy tales, which we tentatively considered the right fit for our purpose, were all that we had on hand. Our new partner needed hundreds of thousands of clearly labelled traing examples for each case of right or wrong behaviour.
We took the texts, chunked them into events (average of 50 in each short tale) and began to tag each event by hand. Now we know that human brain is also chunking narratives into scenes in the same manners as a storyboard artist draws comic-like frames to visualise sequence of events and their internal composition.
We have discovered that even the kindest and the best fairy tale characters didn’t behave according to the playbook of ethics. They frequently committed bad, cruel and hostile actions. Bad characters, on the contrary, sometimes behaved quite consistently in the kindest possible way. They stayed bad at the end anyway. It looked more and more improbable that we could use supervised deep learning successfully for our purpose. The supervised deep learning so far is the only one type of AI that found its way to the real life applications. We just kept going.
I don’t remember the reason why we decided to use tags “magic” and “unexplained” to classify events. It was a gut feeling, I guess, but that was how uncertainty and, in particular, unexpected uncertainty first made its way into our project. In summary of our findings in August 2017 I wrote:
“1. Fairy tales are a language in their own right similar to a sheet music score (The Structural Study of Myth, Claude Levi Strauss): events are letters (notes), their strings are words (melodies).
2. We ‘understand’ the ‘meaning’ of this language implicitly because fairy tales (like all good stories) are subversive not educational (Making Stories, Jerome Bruner).
3. Fairy tales “domesticate uncertainty” (Making Stories, Jerome Bruner) because they use unpredictable events of different scale to connect small everyday life’s unexpected events and great unimaginable miracles.”
We didn’t know then how well our observations fit with recent findings of neuroscience. Now we know that:
“The ability of humans to learn changing reward contingencies implies that they perceive, at a minimum, three levels of uncertainty: risk, which reflects imperfect foresight even after everything is learned; (parameter) estimation uncertainty, i.e., uncertainty about outcome probabilities; and unexpected uncertainty, or sudden changes in the probabilities.” More info…
Later we have found a lot of research material that demonstrates that human brain’s hippocampus and cortical networks mediated by it perform very interesting interrelated processes which form the foundation of human conscious experience:
1. Automatically encode/retrieve several types of uncertainty as a major ground for hippocampal ability to automatically predict future probabilities and make probabilistic decisions.
2. Automatically encode personal real life experiences and important sensory inputs in a number of interrelated formats: (i) as a relational spatial map of landmarks or, in other words, a network of nodes representing major items present in one or more of our life scenes, (ii) as dynamically assembled schemas of scenes and (iii) as narratives dynamically composed of scenes.
3. Automaticall selects the appropriate thinking/learning strategy based on the evaluation of both predictability of the environment and the credibility of data about it.
4. Automatically resolves the exploration/exploitation dilemma based on the unoredicted uncertainty.
We are still in the middle of our quest. If you are interested to follow or to join it, go ahead!
i. Our brain can detect and measure uncertainty as a Bayesian statistician.
Research findings show that the human brain performs better than “classical learning algorithms predict, and indeed makes near-optimal use of all of the available evidence when updating its internal model.”
“Neuronal activity correlated with the predictability of ongoing auditory input, both in terms of deterministic structure and the entropy of random sequences, providing clear neurophysiological evidence of the brain’s capacity to automatically encode high-order statistics in sensory input.”
ii. Brain resolves exploration vs exploitation dilemma through encoding of unexpected uncertainty
“Learning to choose between multiple unknown prospects, in the hope of eventually exploiting the most rewarding ones, is a difficult yet fundamental problem. It involves a trade-off between two competing courses of action: to exploit known options that are believed to yield the best outcomes versus to explore unknown alternatives that may be even more rewarding.”
“Also, the current evidence that unexpected uncertainty induces novelty seeking in the action selection rule, together with prior evidence that unexpected uncertainty plays a key role in value updating (e.g., Behrens et al., 2007 and Payzan-LeNestour and Bossaerts, 2011), suggests that unexpected uncertainty plays a dual role, as a modulator of learning as well as of action selection.” More info…
iii. Adult brain’s ability to detect unexpected uncertainty is blocked by mental models.
“ In our experiment, many participants failed to recognize the presence of unexpected uncertainty. Consequently, in the exit questionnaires they often took the arms to be “random” [in our language, risky] which illustrates the antagonistic relationship between risk and unexpected uncertainty — jumps were confounded with realization of risk.
Our participants’ failure to detect jumps may suggest that their “mental models” excluded nonstationarity a priori.” Almost 90% — 67 out of 75 participants of the experiment failed to detect unexpected uncertainty. More info…
iv. Mental model blocks can be removed by instructions.
“Full Bayesian updating is reflected in human learning only if enough structural information of the outcome generating process is provided. Specifically, the ability to track unexpected uncertainty, and hence, to detect jumps in the outcome probabilities, appeared to rely on instructions that such jumps would occur.” More info…
v. Spatial learning/thinking and response learning/thinking can be used only alternatively. All people spontaneously prefer one of them.
“To reach a target location, one may use a “spatial memory strategy” by learning the relationships between environmental landmarks (stimulus-stimulus associations). This strategy is a form of explicit memory based on a cognitive map… that allows a target to be reached in a direct path from any given direction. This type of flexible navigation has been shown to depend on the hippocampus…. Alternatively, one can navigate without knowledge of the relationships between environmental landmarks, but instead by using a series of turns at precise decision points or stimuli (e.g., turn left at the corner, then turn right after the park, etc.). The successful repetition of this nonspatial strategy leads to a “response strategy” (stimulus–response associations) known to involve the caudate nucleus, a form of implicit memory or habit.”
“Young healthy participants spontaneously use different strategies in a virtual radial maze, an adaptation of a task typically used with rodents. Functional magnetic resonance imaging confirmed previously that people who used spatial memory strategies showed increased activity in the hippocampus, whereas response strategies were associated with activity in the caudate nucleus… Results showed that spatial learners had significantly more gray matter in the hippocampus and less gray matter in the caudate nucleus compared with response learners. Furthermore, the gray matter in the hippocampus was negatively correlated to the gray matter in the caudate nucleus, suggesting a competitive interaction between these two brain areas. ” More info…
vi. Gray matter volume in relevant brain areas correlates with a strategy in use.
“There is a large amount of evidence that supports the hypothesis that the use of spatial strategies is associated with increased hippocampal grey matter and activity, while the use of response strategies is associated with increased grey matter and activity in the striatum (caudate nucleus).” More info…
vii. Adult spatial learners/thinkers don’t see that they are underserved but it can be corrected
“We found that 84% of children reported using a spatial strategy, indicating a clear bias compared to the 47% of young adults who reported using the same strategy. We also observed that 39% of older adults used a spatial strategy.” More info…
But only 10% can detect unexpected uncertainty that triggers spatial learning (see iii).
viii. Why shall we focus on teens?
Many teenagers in mature markets and, especially, in the US are spontaneously rejecting the response learning focused algorithms of modern mass social networking platforms. The growth of Snapchat on exclusive teenage audience (From 173 million active daily users of Snapchat more than 80 million are exclusive — they do not use Facebook, Instagram, WeChat, Youtube or Twitter at all). This data confirms research results, which show that 84% of kids and teens are spontaneous spatial learners/thinkers.
ix. The evidence that Facebook and alike satisfy only response learning/thinking comes from the former top insiders of the social network.
“The short-term, dopamine-driven feedback loops we’ve created are destroying how society works,” said former Facebook executive Chamath Palihapitiya, speaking of Facebook and other social media companies, during the Stanford talk. “No civil discourse, no cooperation; misinformation, mistruth. And it’s not an American problem — this is not about Russians ads. This is a global problem.”
“The company’s former president Sean Parker also recently said the product was engineered to exploit human psychology by providing “a little dopamine hit every once in awhile.” A former early investor and the company’s former privacy chief have also publicly expressed regrets.” Source…
x. Hippocampus plays critical role in flexible cognition and social behavior.
“We also note that the contribution of the hippocampus to flexible cognition is perhaps most apparent in the complex dynamics of social interactions. In everyday social interactions, subtle contextual differences (e.g., a single prior interaction with an individual) require extensive and flexible modifications of our behavior, driving us to select different words, draw upon shared knowledge, or use entirely different language and social conventions for interaction.” More info…
xi. Hippocampus is our internal storyteller and more…
“The right hippocampus is still viewed as encoding spatial relationships, but the left has the altered function of storing relationships between linguistic entities in the form of narratives.” More info…
“Scene construction provides the stage on which the remembered event is played or the ‘where’ for the ‘what’ to occur in… Moreover, we argue that scene construction is an excellent candidate for a common core process that underpins a host of related cognitive functions … including navigation and imagination.” More info…
“In conclusion, we reveal two distinct types of hippocampal representations. We showed that the hippocampus codes for nodal representations (Eichenbaum et al., 1999) where activity patterns represent the “essence” of an item in memory, which is common across different events featuring that item. We also showed that, in addition to item-specific nodal representations within a narrative, the hippocampus also codes for the entire narrative, which may reflect networks of events related through the nodal representations, which feature with relatively different frequencies in each of the two narratives. In combination, the evidence of both item-specific and narrative-specific representations in the human hippocampus suggests that human episodic memories may be subject to hierarchical organization… and may answer the outstanding question of how the brain can simultaneously support seemingly conflicting operations of individuating and flexible recombining of memories … Conceptually, narrative-level representation is similar to other forms of contextual representations, such as temporal and spatial contexts, and similarly to spatial contextual representations, the representations of different narrative contexts diverge over time. The neural mechanisms, which subserve narrative context formation shown here, may be involved in organization of memories of related autobiographical events into personal narratives.” More info…
xii. Model-based Bayesian reinforcement learning.
“We applied a model-based Bayesian learning algorithm… to track subjects’ estimates of the outcome probabilities on each arm. This algorithm provides a principled way to measure unexpected uncertainty, as well as estimation uncertainty and risk, while specifying how they should influence the rate of learning.” More info…
“Our findings, therefore, demonstrate that the human brain has the capacity to disentangle uncertainty into its various components, i.e., risk, estimation uncertainty, or unexpected uncertainty. The resulting signals affect the learning rate differentially and optimally, in line with Bayesian learning.” More info…
xiii. Conscious AI as a by-side product?
“Just as our ancient ancestors may not have been conscious in ways that we can understand, artificial intelligence may soon become conscious in ways that could threaten our race’s existence. Not only does language hold the keys to understanding human consciousness, but it could also allow us to shape machine consciousness in a way that benefits humanity.” More info…
“A human player will initially tell its character what to do like the voice of god from Julian Jaynes’ theory of bicameral minds. The character will be represented by a machine learning algorithm complimented with an analog “I” placed at the centre of a spatial network of item-nodes and, simultaneously, placed into the first person view position in all schemas of scenes-chunks of the narrative of the artificial agent’s (the character’s) life.
The objective of the game will be for the player to train his character to make independent decisions and to act fully autonomously using the spatial representations of its artificial consciousness (network of items, schemas of scenes, narrative of scenes, etc.) and the unconscious probabilistic computations (Shannon entropy, three types of uncertainty, Bayesian model based reinforcement learning, roaming entropy, ets.).”
The above basics don’t cover all our findings but they put you on the same page with us. The rest will be learned in action.
Call to Action
Can we launch a new Renaissance with a ridiculous gaming/content management algorithm that will explicitly entertain people and implicitly change the way how they play, live, buy and die?
Our meta research shows that Western civilization faces the major problem of excessive automation of humans. Because of it:
- young people feel lonely and anxious;
- older people suffer from neurodegenerative diseases;
- our society gets crippled by polarization, anxiety and hate;
- our civilization becomes fragile in the face of change.
Fortunately, neuroscience has recently discovered the root of the problem. Now we propose a plan how to begin fixing it in an economically viable way.
A vast and mounting research in neuroscience demonstrates that the humankind is divided into two tribes by a strategy which each tribe prefers to apply in learning and thinking:
- spatial learners tend to initially associate several stimuli with each other and to link them to a response later to eventually achieve better understanding and bigger rewards. They prefer to explore.They create cognitive maps of physical and abstract spaces which give them more freedom of choice in selecting their paths. It is harder to predict what they do and to control them;
- response learners focus at short stimulus-response feedback loops with quick rewards and learn their paths as sequences of turns. They like to be instructed. Their freedom of choice is limited to learned paths only but they can master their performance to the level of habits which they perform automatically. They are easier to predict and to control.
All healthy people can spontaneously apply either strategy depending on the situation but under the modern, economic, cultural and social conditions the number of people who prefer spatial learning constantly decreases with aging: children — 84%, young adults — 47%, older adults — 39%.
Anatomically spatial learning is associated with hippocampus. Response learning is linked to caudate nucleus. In a relatively stable and predictable environment people tend to apply response strategy more frequently at the expense of spatial strategy. As a result their hippocampus shrinks and their caudate nucleus grows.
In the long run it’s dangerous because shrinkage of hippocampus beyond a particular threshold may lead to Alzheimer’s disease,dementia, Parkinson’s disease.
In the shorter perspective greater caudate nucleus makes people intolerant to uncertainty. Smaller hippocampus that also works as a switch between the two alternative learning strategies gets jammed in the position of response learning. A positive feedback loop emerges that leads to deepening of the problems outlined above.
Furthermore, overdosing of response strategy leads to emotional microdosing because it develops habits which don’t require nor produce emotional rewards for their execution. Lowered level of emotions squanders human ability to make decisions, including buying decisions among others, because our brains need emotions to trigger decision making processes — bigger the decision, stronger the emotions are required for it to happen.
To further reinforce this vicious circle, economical, cultural and social environments are generating mental models which develop aversion of any uncertainty and block natural ability of 90% of adult people to differentiate uncertainty, in particular, to detect unexpected uncertainty that should, under normal conditions, initiate spatial strategy based explorative behavior. This behavior, which some scientists call roaming entropy, produces a regenerative impact on hippocampus by launching a successful neurogenesis process in it.
Spatial learners enjoy unexpected uncertainty and roaming entropy because of their resulting stronger emotions. The vast majority of currently available entertainment and media, on the contrary, is designed for response learners. Spatial learners can consume this product but it leaves them unsatisfied. This discrepancy between the need and the supply creates a unique window of opportunity.
To capture this opportunity we develop a game, or rather a gaming platform that will incorporate unexpected uncertainty, roaming entropy and reinforcement learning as ridiculous algorithms which will allow us to:
- entertain spontaneous spatial learners (84% of kids and teens, 47% of young adults, 39% of older adults),
- help spontaneous response learners to restore their spatial learning capacity and rejuvenate their hippocampus,
- prevent and, in some cases, revert hippocampus linked neurodegenerative disorders,
- provide to online retailers new highly effective tools and a platform to increase sales to spontaneous spatial learners.
The cultural movement of the scale of Renaissance is, absolutely, required to successfully dismantle mental models which block spatial learning in human heads on a global scale. Indeed, it will be another Renaissance because along with resolving problems and eliminating threats it will open unparalleled new opportunities. Reactivation of spatial learning is a mandatory prerequisite of exponential growth of human self learning capacity. The vast self learning capacity, in its own turn, is absolutely necessary to make the enormous knowledge accumulated by humanity equally beneficial for all. Self learning, self thinking, self reliant individuality should thus become the next grand narrative of the human civilisation. With our ridiculous algorithms we are only taking the first small step in this direction. Together we can do more.