A box-and-arrow overview of a learning software intelligence, with unconsciousness and emotions, capable of defending itself when under threat. The architecture combines recent and prominent models from cognitive psychology, developmental psychology, logic, linguistics, social psychology, and psychoanalysis.
While there has been much prior criticism of the possibility of modeling cognition (Weizenbaum, 1976; Nisbet, 1977), Chalmers (2011) recently argued that sufficient information is now available to model cognition to the level of consciousness itself, which may not be ultimately knowable but probably exists (Nagel 1974). While being an interesting debate as to whether such a machine would indeed by conscious, the intent of this project is simply to integrate existing knowledge in the fields of cognitive psychology, developmental psychology, linguistics, logic, social psychology, and psychoanalysis to create a viable model for a self-aware software intelligence. Most AI systems at this level attempt to create an existing sophistication, as if we are born fully cognisant. The approach here instead is to consider a learning AI, such as HAL in the famous Kubrick movie, "2001: A Space Odyssey."
As basis, therefore, this system prefers Demetriou's recent 'hypercognitive' model of mental development (Demetriou, 2006; Demetriou and Spanoudis, 2006). Demetriou defines three stages of processing in cognition (control, domain, and executive) with six domains (categorical, numeric, causal, spatial, social, and linguistic), developed from tabula rasa. This is the modern extension of Piaget's theory of development from 'blank slate' (1925), modified to account for anomalies in the original formulation, particularly adherents of Vygotsky (1926) and many corroborations (for example, Kozulin et al, 2003). The 'blank slate' model for empty cognition at birth itself dates to Aristotle (ca. 400 BC) and was,moreover, the basis of the ethical and legal system used for the definition of the United States as a nation, that people are 'born equal' (derived from Locke, 1690). So it is not a new idea, although recently, many neurologists have challenged this view (notably Pinker, 2002).
This system takes the balanced approach between a blank slate and a totally preprogrammed neocortex (as suggested by Dupre, 2002), including elements necessary to avoid the Nisbett's introspection illusion (1977), and moreover, by doing so,allowing for judgment distortion caused by peer and authority pressure (Asch, 1951; Milgram, 1961; Stanford, 1971; Haidt, 2000; Prasad et al., 2009; Nisbet and Garrett, 2010).
1.1. Prior Topics on this Site
This particular topic only provides a summary of the model.The following topics provide background discussion on how Demetriou's box-and-arrow model can be extended as the basis for a software intelligence:
- "Designing HAL (1): Hypercognition, Freud and Jung" showed that the depth of Demetriou's model permits inclusion of the model of the psyche, proposed in different forms by Freud (1933, 1940) and Jung (1935, 1971). Jung's model was also found commensurate with modern ideas of cognitive dissonance (Festinger, 1957), which is the accepted explanation of social pressure causing cognitive distortion, and to provide a basis for unconscious social archetypes in this model.
- "Designing HAL (2): A Learning AI with Emotions and Dreams" outlined a framework for designing a complete computer model of cognition, including emotions and dreams, using simulated annealing with perturbation as a model to explain how information is reorganized during dreams.
- "Designing HAL (3): Developing Emotions and Self Awareness" described the basis for adding symbols to the cognitive network, explaining the indirect creation of self perception from intuitive emotions, and from that, self awareness.
- "Designing HAL (4): Play and Archetypal Formation in Categorical Processing Domain" defined the processing necessary in the categorical domain, outlining the initial archetypes and introducing how their development would affect decision making.
Asixth part of this series (not published on this site) compares various software tools for writing the system. This includes:
- Neural-network modeling systems for definition of relative abstractions, and how well they can be integrated with compiled simulated annealing code for dreams and play (which is a non-trivial task, as it requires thread balancing with inter-process messaging and semaphores between object code generated by different compilers).
- A GUI to a human 'language teacher' who would associate the generated symbols with natural-language correlations (probably as pictograms, for each of demonstration).
- A training program, which would provide various stimuli and 'food' on a time schedule for building up object sets, without requiring undue human interaction.
The remainder of this topic describes the objects, properties, and events at an architectural level.
2. Active Memory
In current cognitive psychology, 'consciousness' is thought to be a property of the 'executive function,' first proposed by Broadbent (1987), and refined by Shiffrin and Posner (1975). The input and results of 'conscious' decision making pass into active, or 'short-term' memory, which provides three separate inputs into the control stage (Raaijmakers, 1993; Baddeley, 1996; Nuxall, 2007; Cowan, 2010).
In most representations, active memory is considered a portion of the executive function, but the division is somewhat arbitrary, and for the purposes of this model, the active memory is shown at the same level. In this model, the contents of active memory are also available to decision making for comparison. However, the process of committing active memory to long-term memory, which is known to take many seconds, is separate, so in this system, that which is denoted as 'short-term' memory is really just a buffer for temporary storage of information delivered from the processing domains, together with current decisions, for adding to the processing domains through the control stage.
In accordance with current theory, the decision-making process has access to the three kinds of information stored in short-term memory. If the decision-making process is not called upon to make new decisions in conflict with information in the same domain as that in active memory, a new symbol arising from the executive function can be added to the domain-processing stage by a completed act of introspection.
3. Symbols as Multi-domain Objects
In traditional artificial-intelligence software, one processing algorithm is applied to complex data sets. Research in developmental and cognitive psychology suggest that multiple processes occur simultaneously in cognition, however. Demetriou's model provides for up to six separate kinds of processing that occur in parallel on data in different domains, which then is passed up to a higher-level system for decision making and motor control.
The proposed extended model defines a semiotic process of adding symbols to the processing domains as the result of new cognitions. It further extends the Demetriou model by adding a 'depth' parameter to the symbols (defining how accessible they are to conscious processing in the executive function), as well as an intuitive emotional complex (providing unconscious motivation and possibly inhibiting conscious control).
The degree of access to the symbols by the executive function is set by the global 'wakefulness' parameter. Wakefulness is a symbol whose primary value is set from the physiological state by the control system. When wakefulness is higher, the executive function can access more symbols; when it is lower, the executive function can access less symbols.
3.1. Symbol Creation
There are up to six sets of properties attached to each symbol, and six kinds of processing applied to those symbols during the cognition process. The symbol also contains a 'scoreboard' with indication of its successfulness (via a 'utility' parameter) and depth. Each time a cognition results in success the utility value is incremented. When the utility value of two connected symbols exceeds a threshold:
- A new symbol is created.
- The depth of the symbol is decreased, and the old symbols' depths are increased, effectively inserting the symbol between the lower symbols and the executive function
- Success points are transferred to the new symbol.
- Other values for the new symbol are acquired for the new symbol from the original symbols.
3.2. Symbol Properties for Multi-domain Processing
Each symbol has one global and six dedicates sets of properties:
- Global Control - Depth, utility, domain pertinence
- Categorical - Connection (parent/child/peer) and amount (Salii, 1965; Kuzmin, 1982; Cornelis, 2003)
- Numerical - Number (inapplicable/0-x/many), type (collection, sequence)
- Causal - Level (1st/2nd/3rd order logic), type (order dependent)
- Spatial - Abstraction of symbolic results from lower-order multidimensional processing
- Social - Attribution of agents (involuntary/voluntary) to archetypes (peer, maternal, paternal)
- Linguistic - Mapping of abstract symbols to words and syntactical connection
3.3. Domain Pertinence
With a full list of symbols, an additional six control variables required to indicate the pertinence of domain processing assumed for each symbol.
3.4. Symbol Reorganization
The system allows for reorganization and reassignment by tow methods:
- Through dreaming, background process, modeled on simulated annealing with perturbation, as described in the second article in this series (Henderson et al., 2002)
- Through play, described in the fourth article (Breazal 2005, 2006).
3.5. Deferred Development of Social and Language Skills
Social and Linguistic development are deferred. While elements of the social processing domain are believed necessary to be present at birth, full development of social domain does not occur until after puberty.
As linguistic development is also deferred, the relation of linguistic symbols to primary symbols in the infant cognition is difficult to state, as the symbols have already become large complexes by the time they are mapped to words. The model for linguistic development in the HAL system is based on the concept of universal language (Chomsky, 1965), although there are many critics, most recently Sampson (2005), it is amenable to computer modeling.
4. Emotions and Archetypes
The model assumes a standard set of six basic emotions (Cannon, 1915; Ekman, 1999; Breazeal, ca. 2002):
- Fight/flight (also known as anger/fear)
- Love/hate (or like/dislike)
The emotions are abstracted to form archetypal models for voluntary agents (peers, maternal agents, and paternal agents).
4.2. Congruence and Dissonance
The Jungian model of the psyche proposes three main elements: the persona, self, and shadow (Jung, 1935, 1971). The rational self projects a persona onto the world, based on rationalization of underlying archetypal representations of individuals in the unconscious shadow by the rational self. When there are irrational fractures in the unconscious that cannot be resolved, the individual cannot help but display unconscious behavior that may be to the detriment of the person. An example is an individual displaying disgust and insulting behavior when confronting an opposing view in rational argument.
The Jungian model of the psyche is analogous with current theories of 'cognitive dissonance,' permitting the integration of psychoanalytical observations into the cognitive model.
In this model, the emotions therefore control unconscious actions, which in the cognition system are passed into the executive function unconsciously. The act of introspection forms an abstracted perception of self, which passes into the motivation section of the executive function. If the perception of self is in dissonance with the instinctive emotional response, the motivation is inhibited, and the result of decision making in the executive function is uncertain,leading to further introspection. If the intuitive emotional response is in congruence with the perception of self, the decision making process is more certain, and the utility of the symbols leading to the cognition is increased.
4.3. Refinement of Self Perception
Intuitive emotions function independently of cognition, creating new symbols independent of introspection. Introspection on the nature of one's emotions leads to the creation of new symbols representing the self, which are separate from the intuitive emotional complex and result in the formation of the self perception passed to the motivation center of the executive function.
Introspection on the perception of self creates another, separate cluster of symbols which are passed to the decision-making process as the conscious self identity.
If introspection on the self leads to certainty in decision making, a second inhibition path from the decision making process provides conscious inhibition of autonomous motor actions.
5. Decision Making
Motivation is derived from emotions and physical needs in a hierarchy (Maslow, 1943). Decisions are affected by emotion (Weizenbaum, 1976; Kiefer, 2012). New symbols are created by the executive function posing 'what-if' scenarios and examining the available representations, categorizing them, and using them to form a deduction.
6. Language DevelopmentBecause the language processing domain is separate from the domains where symbols representing observations and internal state are created, words themselves do not have intrinsic meaning (as in Plato, ca. 380 BC) bur rather are formed by a causal reference (Kripke, 1980). Initial language is formed by connecting objects, actions, and events with specific words (Davidson et al, 1957; Davidson, 2001) and organized by a basic syntax supporting up to three words, in accordance with the theories of universal grammar (Chomsky, 1965). Initially there is no cognition as to how the words connect to grammatical elements; the universal grammar simply defines all possible combinations of grammatical elements which may be assigned to a sequence of words. In later development, the significance of underlying principles in grammatical constructions allows the understanding and creation of more sophisticated word complexes for communication.
This is an aggregate list of references for the series of topics on hypercognition.
- Aristotle (ca. 360BC)On the Soul
- Asch conformity experiments:
- Baddeley, A; Sala, S. D. (1996) "Working Memory and Executive Control," Philosophical Transactions: Biological Sciences 351.1346, Executive and Cognitive Functions of the Prefrontal Cortex
- Breazeal, C. "Sociable Machines" (ca. 2002)http://www.ai.mit.edu/projects/sociable/emotions.html
- Breazeal, C.; Gray, J.; Berlin, M. (2006) "An Embodied Cognition Approach to Mindreading Skills for Socially Intelligent Robots," MIT
- Breazeal, C.; Thomaz, A. (2005) "Learning from Human Teachers with Socially Guided Exploration," MIT
- Broadbent, D.E. (1987)Perception and Communication, Oxford University Press
- Cannon, W.B. (1915)Bodily Changes in Pain, Hunger, Fear and Rage: An Account of Recent Researches into the Function of Emotional Excitement, Appleton.
- Chomsky, N. (1965)Aspects of the Theory of Syntax, MIT Press.
- Cognitive Dissonance
- Cornelis, C.; De Cock, M.; Kerre, E. "Intuitionistic fuzzy rough sets: at the crossroads of imperfect knowledge," Expert Systems, v. 20, issue 5, pp. 260–270, 2003
- Cowan,N. (2010) Multiple Concurrent Thoughts: The Meaning and Developmental Neuropsychology of Working Memory, Dev Neuropsychol.
- Davidson, D.; Suppes, P.; Siegel, S. (1957)Decision-Making: An Experimental Approach, Stanford University Press.
- Davidson, D (2001),Essays on Actions and Events, Oxford University Press.
- Demetriou, A. (2006)Neo-Plagetian Theories of Cognitive Development.
- Demetriou, A ; Spanoudis, G.; Mouyi, A. (2006) "A Three-level Model of the Developing Mind: Functional and Neuronal Substantiation"
- Dupre, J. (2002) "Making Hay with Straw Men" American Scientist
- Ekman, P. (1999). "Basic Emotions". In: T. Dalgleish and M. Power (Eds.). Handbook of Cognition and Emotion. John Wiley & Sons Ltd, Sussex, UK:.http://web.archive.org/web/20101228085345/http://www.paulekman.com/wp-content/uploads/2009/02/Basic-Emotions.pdf
- Festinger, L. (1957).A Theory of Cognitive Dissonance.California: Stanford University Press.
- Freud, S. (1933) New Introductory Lectures on Psychoanalysis, Penguin Freud Library
- Freud, S. (1940) An Outline of Psycho-analysis
- Haidt, J. (2000) "The Emotional Dog and its Rational Tail: A Social Intuitionist Approach to Moral Judgment," University of Virginia
- Henderson, D.; Jacobson, S.; Johnson, A. (2002) "The Theory and Practice of Simulated Annealing."
- Jung, C. G. (1935)Structure & Dynamics of the Psyche,The Collected Works of C. G. Jung, Volume 8. Princeton University Press
- Jung, C.G. (1971).Psychological Types, Collected Works, Volume 6, Princeton, N.J.: Princeton University Press
- Kiefer, M. (2012) "Executive control over unconscious cognition: attentional sensitization of unconscious information processing," Frontiers in Human Neuroscience, 6, 61.
- Kozoulin, A. (ed.) et al (2003)Vygotsky's Educational Theory in Cultural Context, Cambridge University Press
- Kripke, S. (1980)Naming and Necessity. Harvard University Press:
- Kruger and Dunning (2006) "Unskilled and Unaware of It: How Difficulties in Recognizing One's Own Incompetence Lead to Inflated Self-Assessments," Journal of Personality and Social Psychology, 77 (6): 1121–34. doi:10.1037/0022-3522.214.171.1241
- Kuzmin,V.B. (1982)Building Group Decisions in Spaces of Strict and Fuzzy Binary Relations, Nauka, Moscow, 1982 (in Russian)
- Locke, J. (ca. 1690)Essay Concerning Human Understanding, Volume 2
- Maslow, A.H. (1943) "A theory of human motivation," Psychological Review 50 (4) 370–96
- Milgram experiment:
- Nagel, T. (1974) "What Is It Like to Be a Bat?" The Philosophical Review, Vol. 83, No. 4 (Oct., 1974), pp. 435-450
- Nisbet; E.; Garrett K. (2008) "Belief in rumors Hard to Dispel: Fact checking easily undermined by images, unrelated facts," Ohio State University
- Nisbett, R.E.; Wilson, T.D. (1977) "Telling more than we can know: Verbal reports on mental processes".
- Nuxoll, A. (2007) "Enhancing Intelligent Agents with Episodic Memory"
- Piaget, J. (1925)The Child's Conception of the World, London: Routledge and Kegan Paul.
- Pinker, S. (2002)The Blank Slate: The Modern Denial of Human Nature. Continuum
- Plato (380 BC)The Republic, Allegory of the Cave, 514a–520a
- Posner, M.I.; Snyder, C.R.R. (1975) "Attention and Cognitive Control," Information processing and cognition: the Loyola symposium. Hillsdale, N.J: L. Erlbaum Associates.
- Prasad, M., Perrin, A; Bezila, K; Hoffman, S.; Kindleberge, K; Manturuk K; Powers, A. (2009) "There Must Be a Reason”: Osama, Saddam, and Inferred Justification," Sociological Inquiry, 29.2, pages 142–162
- Raaijmakers, J.G (1993) "The story of the two-store model of memory: past criticisms, current status, and future directions". Attention and performance. XIV (silver jubilee volume). Cambridge, MA: MIT Press
- Salii, V.N. (1965)Binary L-relations, Izv. Vysh. Uchebn. Zaved., Matematika, v. 44, No.1, pp. 133–145
- Sampson, G. (2005)Educating Eve: The 'Language Instinct' Debate, continuum.
- Shiffrin, R.M.; Schneider, W. (March 1977) "Controlled and automatic human information processing: II: Perceptual learning, automatic attending, and a general theory," Psychological Review
- Stanford Prison Experiment:
- Thorndike E.L.; Woodworth R. (1901) "The Influence of Improvement in one Mental Function upon the Efficiency of Other Functions," Psychological Review, 8, 247-261.
- Vygotsky, A.Educational Psychology, 1926.
- Weizenbaum, J. (1976)Computer Power and Human Reason: From Judgment To Calculation,San Francisco: W. H. Freeman