why metacognition and social cognition will become critical … · 2019. 5. 23. · •solutions...
TRANSCRIPT
Jerald D. KralikDepartment of Bio & Brain Engineering
KAIST
Why Metacognition and Social Cognition
will become critical components of AI
Jerald D. KralikDepartment of Bio & Brain Engineering
KAIST
Why the (Human) Mind/Brain, Metacognition and
Social Cognition are so important for AI
Why (Human) Mind/Brain for AI?
• Important work not using it as model
– More purely derived from problem being solved
– But good luck
• Why not use it?
– Does what most of us are trying to do/accomplish
– Both expert and artificial general intelligence
• A suboptimal kluge
– Mostly myth
Why Study the Mind/Brain?
• Extraordinary system/device that works
• Solutions provide further benefits
– Insight to inform our choices
• Natural tradeoffs
• Suboptimal from different perspectives
– Health/Dysfunction
What Does the Brain Tell Us?
• We do know a great deal
• Yet still determining principles, general structure (and details)
• Especially higher cognitive abilities
• So: cognitive and brain science a challenge
Why Not Farther at this Point?
• Missing overarching, principled theoretical approach
– Lacking a theory of brain function
• Even with the most fundamental units—neurons
– What are they trying to do?
• And at the top: the purpose/structure of high-level
function
– A theory of human intelligence
• Or even a consensus re: general cognitive architecture
NMSNST
MCN output
High-level cognition in the primate brain
basal
ganglia
posterior
parietal
S1/S2
V1-V5
IT
basal
forebrain
Amyg Hipp
M1PM
PFdl
PFm
OFC
Thal
Hθ
Anterior Posterior
Need Theoretical Framework
A Standard Model of the Mind: Toward a Common Computational
Framework Across Artificial Intelligence, Cognitive Science,
Neuroscience, and Robotics.
JE Laird, C Lebiere, PS Rosenbloom - Ai Magazine, 2017From Wikipedia
My Approach
• Literature review from relevant fields
– Evolutionary biology
– Psychology
– Neuroscience
– Anthropology
– Sociology
• Evolutionary/Anthropological Approach
• Logical analysis
Kralik, J. D., Mao, T., Zhao, C., Nguyen, H. T., and Ray, L. E. (2016). Modeling incubation and restructuring for creative problem solving in robots. Robotics and Autonomous Systems, Special Issue on Robotics and Creativity, 86: 162-173.
Kralik, J. D., Shi, D., El-Shroa, O. A., and Ray, L. E. (2016). From low to high cognition: A multi-level model of behavioral control in the primate brain. Proceedings of the Annual Meeting of the Cognitive Science Society. Selected for oral presentation.
Kralik, J. D. (2017). Architectural design of mind & brain from an evolutionary perspective. Proceedings of the AAAI 2017 Fall Symposium: A Standard Model of the Mind.
Kralik, J. D., Muldrew, D. B. C., Gunasekaran, D., and Lange, R. D. (2017). Cognitive control for goal-directed reaching in a humanoid robot. Proceedings of the IEEE International Conference on Robotics and Biomimetics (ROBIO). Selected for oral presentation.
Ray, T. H. and Kralik, J. D. (2017). Seeking true intelligence from the ground up: Evolutionary origins of cognition. Proceedings of the IEEE International Conference on Robotics and Biomimetics (ROBIO). Selected for oral presentation and session chair/moderator (Kralik).
Kralik, J. D. (2018). Core High-Level Cognitive Abilities Derived from Hunter-Gatherer Shelter Building. In I. Juvina, J. Houpt, & C. Myers (Eds.), Proceedings of the 16th International Conference on Cognitive Modeling (pp. 49-54). Madison, WI: University of Wisconsin.
Kralik, J. D., Lee, J. H., Rosenbloom, P. S., Jackson, Jr., P. C., Epstein, S. L., Romero, O. J. , Sanz, R., Larue, O., Schmidtke, H. R., Lee, S. W., McGreggor, K. (2018). Metacognition for a Common Model of Cognition. Procedia Computer Science: 145, 730–739.
Lee, J., Kralik, J. D.*, and Jeong, J.* (2018). A Sociocognitive-Neuroeconomic Model of Social Information Communication: To Speak Directly or To Gossip. Proceedings of the Annual Meeting of the Cognitive Science Society. *Co-corresponding authors
Lee, J., Kralik, J. D.*, and Jeong, J.* (2018). A General Architecture for Social Intelligence in the Human Mind and Brain. Procedia Computer Science: 145, 747–756. *Co-corresponding authors
References
–Associative Learning II
–Associative Learning I
– Innate Systems
– Causal Reasoning I
– Causal Reasoning II
– Reductionism I
– Reductionism II
– Reductionism III
– Reductionism IV
Co
ntro
l Circu
it Lev
els
– Defense– Mating– Social
– Foraging & Ingestion
– Physical Environment
Action Control Circuit
{Stimuli}
Content Domains
Type 4:Abstract,
Reasoning-Based
Type 2:Content-Specific,
Associative-Based
Type 1:Content-Specific,
Innate Systems
Type 3:Content-Specific,
Reasoning-Based
Type 4:Arbitration (Control & Monitoring)
Cortex & Basal Ganglia (Nac-Shell)
Hypothalamus/Midbrain Complex
Cortex & Basal Ganglia (Nac-Core)
Neocortex: Agranular PFC
1st Granular PFC, PPC, TC
2nd Order PFC, PPC, TC
3rd Order PFC, PPC, TC
4th Order PFC, PPC, TC
5th Order PFC, PPC, TC
ACC, Lateral PFC
Type 5:Metacognition
(Monitoring, Modulation & Control)
Error
Long-
Term
Memory
Representation Decision-Making Action{SEI} Outcome
Restructuring
Affective/
Goal
Gate
Feedback
(for Learning)
Example Representation: A Grid World
Initial State
Goal State
* Gray squares are obstacles
We Evolved to Act
–Associative Learning II
–Associative Learning I
– Innate Systems
– Causal Reasoning I
– Causal Reasoning II
– Reductionism I
– Reductionism II
– Reductionism III
– Reductionism IV
Co
ntro
l Circu
it Lev
els
– Defense– Mating– Social
– Foraging & Ingestion
– Physical Environment
Action Control Circuit
{Stimuli}
Content Domains
Type 4:Abstract,
Reasoning-Based
Type 2:Content-Specific,
Associative-Based
Type 1:Content-Specific,
Innate Systems
Type 3:Content-Specific,
Reasoning-Based
Type 4:Arbitration (Control & Monitoring)
Cortex & Basal Ganglia (Nac-Shell)
Hypothalamus/Midbrain Complex
Cortex & Basal Ganglia (Nac-Core)
Neocortex: Agranular PFC
1st Granular PFC, PPC, TC
2nd Order PFC, PPC, TC
3rd Order PFC, PPC, TC
4th Order PFC, PPC, TC
5th Order PFC, PPC, TC
ACC, Lateral PFC
Type 5:Metacognition
(Monitoring, Modulation & Control)
Cross branches for beam reinforcements (walls)
Larger branches (boughs) as foundation beams
Smaller branches
Inserted in ground holes
Twine to tie down structural elementsLeaves or Grass exterior
Ground cleared of brush & debris
Entrance
Simple…right?
Core High-Level Cognitive Abilities Derived from
Hunter-Gatherer Shelter Building
Core High-Level Cognitive Abilities
Key Features that Pop
• Metacognition
• Social cognition
• Emotion
• Causality
• Content-specific (expert) vs Content-free (general)
• More
Key Features that Pop
• Metacognition
• Social cognition
• Emotion
• Causality
• Content-specific (expert) vs Content-free (general)
• More
Jerald D. Kralik1, Jee Hang Lee2, Paul S. Rosenbloom3, Philip C.
Jackson, Jr.4, Susan L. Epstein5, Oscar J. Romero6 , Ricardo
Sanz7 , Othalia Larue8,
Hedda R. Schmidtke9, Sang Wan Lee1,2, Keith McGreggor10
1Department of Bio and Brain Engineering, Korea Advanced Institute of Science and Technology
(KAIST), Daejeon, 34141, South Korea2KI for Health Science and Technology, Korea Advanced Institute of Science and Technology
(KAIST), Daejeon, 34141, South Korea3Institute for Creative Technologies & Department of Computer Science, University of Southern
California, Los Angeles, CA, USA4TalaMind LLC, PMB #363, 55 E. Long Lake Rd., Troy, MI, USA
5Department of Computer Science, Hunter College and The Graduate Center of the City
University of New York, New York, NY, USA6Department of Machine Learning, Carnegie Mellon University, Pittsburgh, PA, USA
7Autonomous Systems Laboratory, Universidad Politécnica de Madrid, Spain8Department of Psychology, Wright State University, Dayton, Ohio, USA
9Department of Geography, University of Oregon, Eugene, OR, USA10College of Computing, Georgia Institute of Technology, Atlanta, Georgia, USA
Metacognition Working Group Members
Metacognition — Why? • Metacognition:
– cognition about cognition
• Addresses what the system knows
• The importance of what is known
• Including what has been remembered and what is worth
remembering or forgetting
• Advantages include
– Management/Orchestration/Arbitration of competing or
complementary functions
– Modulation to help finetune other cognitive processes
– Safeguards against confusion and errors from lower cognitive
processes
• especially those designed for efficiency and specialization
– Data management to reduce inefficiencies
• e.g., removal of obsolete information by forgetting
• A process is metacognitive if and only if:
• it receives input from
• sends output to
• or both receives from and sends to the same process type
• A process type is perception, decision, or action
• Cognition captures the entire perception-decision-action cycle
• Decision is broadly construed
• Including, for example: Reasoning, Planning, Attention
Metacognition Defined
• A process is metacognitive if and only if:
• it receives input from
• sends output to
• or both receives from and sends to the same process type
• A process type is perception, decision, or action
• Decision is broadly construed
• Including, for example: Reasoning, Planning, Attention
Metacognition Defined
DecisionDecision
Decision
• Category 0: cognition itself, with primary input from perception and primary output to action control
• Categories 1, 2, and 3 comprise metacognition
• Category 1: Primary input and output from and to decision processes
• Category 2: Primary input from decision processes; primary output feeds forward to action control
• Category 3: Input from perception; primary output to decision processes
General Categories of Metacognition
Examples of Each Category
• Category 1:
– Arbitration of Model-Free vs Model-Based Reinforcement Learning
– Self-Representation
– Reflection and Self Improvement
– Self-control
• Category 2:
– Social cognition
• Category 3:
– Context and abstract task relevant information
• Organizing
• Maintaining
• Allocating (cognitive resources)
• Regulating
• Modulating
• Modifying
• Replacing
• Configuring (e.g., parameters, goals,
reward functions)
• Healing
• Orchestrating
• Coordinating
• Arbitrating
• Broadcasting
• Recruiting
Components of Metacognition
• Understanding
• Awareness
• Reflection
• Explaining
• Debugging
• Generating
• Adapting
Thinking General Management (of single processes) Control of multiple processes
Eventually Consciousness (?)
• Artificial consciousness
• Emotions too!
Conclusions about Metacognition
• Indeed is occurring organically
• With multiple current approaches
• Mind/brain analysis shows need for explicit understanding and
framework
– Plus provides specifics to help
guide, anticipate
Why social per se?
Why not just smarter, to take over and offline?
• Answer:
– As intelligence of things increases, trying to do more for us will require:
• Anticipating needs
• More interactive assistance
• Examples:
– Personal assistant
– Home
– Work
– Integration with us
• We need to help it as well
• Dynamic, complex, novel activities
Social AI: Theoretically
• F1(Person, F2(AI, Goal))
• F2 = F2.1 & F2.2
• F1(Person, F2.1(AI-affordance, F2.2(AI-action, Goal)))
• AI-affordance = interface with us properly
• And yet we and our world are dynamic and complex!
• Therefore, as artificial intelligence and its prevalence (like IoT)
increases sociality increases
• Or put differently, if sociality capacity of system increases,
value, functionality increases
Why is Sociality so Hard?
Beliefs/Knowledge
Preferences
Goals
Rules/norms
Intentions (action policy)
Plus:
• Social rules, conventions
• Perspective-taking
• Culture
Beliefs/ Knowledge
Preferences
Goals
Rules/norms
Intentions (action policy)
Mind-reading, Perspective-taking, Theory of Mind
Beliefs/Knowledge
Preferences
Goals
Rules/norms
Intentions (action policy)
Plus:
• Social rules, conventions
• Perspective-taking
• Culture
Beliefs/ Knowledge
Preferences
Goals
Rules/norms
Intentions (action policy)
Our Social Intelligence Model
Figure 4: The complete model of social information communication. Brain region abbreviations: fusiform gyrus (FG), posterior superior temporal sulcus
(pSTS), medial prefrontal cortex (MPFC), temporoparietal junction (TPJ), amygdala (AMG), insula, right temporal lobe (rTL), interparietal sulcus (IPS),
anterior paracingulate cortex (aPCC), medial precuneus (med. precuneus), anterior insula (ant. insula), anterior cingulate cortex (ACC), orbitofrontal cortex
(OFC), ventral premotor cortex (vPMC), dorsolateral prefrontal cortex (dlPFC), posterior parietal cortex (PPC), posterior superior temporal cortex (pSTC),
caudate nucleus (CD), ventral striatum (VS), mesolimbic dopamine (DA) system, ventral tegmental area (VTA), premotor cortex (PMC), basal ganglia (BG).
Our Social Intelligence Model v2.0
• With Significant metacognition organically
• Lee, J., Kralik, J. D.*, and Jeong, J.* (2018). A General Architecture for Social Intelligence in the Human Mind and Brain. Procedia Computer Science: 145, 747–756. *Co-corresponding authors
. . .
P
C V
PC V
PC V
Actual
Model
PC V
Individually driven factor
Political
Factor P
C VP
C V
Shorter-Term
Aspect
Mood
General Cognitive Infrastruct
ure
Mind/Brain
Architecture
PC V
Cultural/Societal Factor
PC V
Social Contracts
Moral/Social
Dimension
A1-A2 A2-
A1
A1-A3
A3-A1
...Care/Har
m
General Social
...
...
CareHarm
... ...
Longer-Term
Aspect
Personal backgro
und
General affect,
Cognitive ability and
balance
Model-self:Content domain
Mating/Sexual
Social
Individually driven
Political
Cultural/
Societal
Social Contrac
ts
Moral/Social
Dimension
A2-A1
A1-A3
A3-A1
...Care/Harm
General
Social
...
...
Care
Harm
... ...
A1-A2
Self
P
C V
PC V
PC V
SelfOthe
r
Self-Actual-Model (SAM)
Self representation tree (SAM tree) Sel
fOther
Self
Other
AReceiv
er (A2)ATarget
(A3)
. . .
. . .
. . .
Self-Actual-Model (SAM)
Tree
Model of Mind Tree
PC V
PC V
PC V
• Self & Others
• Self:
– Actual
– Model of Self
• Others:
– Receiver AR
• AR self & others
– Target AT
• AT self & others
Models of Minds
1. For relationship type (ingroup, outgroup, celebrity) more gossiping should
spread about ingroup compared to outgroup, and in some cases ingroup over
celebrities
2. Status effects: certain types of scenarios should generate more gossiping
about celebrities as compared to the other groups
3. Positive scenarios will show gossip spreading rates more comparable to
negative ones
4. Greater spreading of positively valenced scenarios with ingroup targets
5. Greater spreading of negativity about celebrities
6. Negativity should be reduced for ingroup targets
7. Dimensions of morality are predicted to generate the most gossip, especially
those involving more egregious threats, like harm and cheating
8. Differences should be found among the morality domains themselves
9. Even simple social activities should generate higher spreading rates for
ingroup targets, in order to maintain accurate detailed knowledge about
ingroup members
9 Predictions about ‘Gossip’
. . .
P
C V
PC V
PC V
Actual
Model
PC V
Individually driven factor
Political
Factor P
C VP
C V
Shorter-Term
Aspect
Mood
General Cognitive Infrastruct
ure
Mind/Brain
Architecture
PC V
Cultural/Societal Factor
PC V
Social Contracts
Moral/Social
Dimension
A1-A2 A2-
A1
A1-A3
A3-A1
...Care/Har
m
General Social
...
...
CareHarm
... ...
Longer-Term
Aspect
Personal backgro
und
General affect,
Cognitive ability and
balance
Model-self:Content domain
Mating/Sexual
Social
Individually driven
Political
Cultural/
Societal
Social Contrac
ts
Moral/Social
Dimension
A2-A1
A1-A3
A3-A1
...Care/Harm
General
Social
...
...
Care
Harm
... ...
A1-A2
Self
P
C V
PC V
PC V
SelfOthe
r
Self-Actual-Model (SAM)
Self representation tree (SAM tree) Sel
fOther
Self
Other
AReceiv
er (A2)ATarget
(A3)
. . .
. . .
. . .
Self-Actual-Model (SAM)
Tree
Model of Mind Tree
PC V
PC V
PC V
• How scale?
– Family Strangers
• Accuracy
– Genes
– Personal history
– Identity
– Society
• Big data & AI
– Privacy
Issues
Social Cognition Conclusions
• As intelligence increases, need for sociality increases
• Reading minds
– Complex, dynamic, enigmatic (hidden, uncertain)
• Thus, why it is needed
Conclusions
• Bottom-up task driven approaches are fine, valuable
• But lacking overarching theoretical understanding
– to capture complete system as whole
• Theoretical examination of the brain — especially high-level
human abilities — critical for AI
• Exposes and explicates critical abilities
– social and meta cognition
Thank you for this nice social
interaction!