IV Flashcards
Main aspects of Cognitivist modelling approaches
inferencing mechanisms, representations, rationality, problem solving
Architectures of cognitivist approach
SOAR, EPIC, ACT-R, …
Problems of cognitivist modelling approach
Grounding Problem: Connection to the environment
Main aspects emergent modelling approach
robust routine behaviour, understanding and exploitation of coupling to the environment
Problem of Emergent Modelling approach
Autonomy problem - separation from one’s environment
How should/shouldn’t we obtain a unified theory of cognition?
Not: divide-and-conquer - highly specific theories of very limited range of phenomena
But: Focus on integrative work= cognitive architectures
= include attention, memory, problem solving, decision making, learning
Typical Application area of unified theories of cognition
human-machine interaction
A wide range of cognitive capabilities are employed in “even simple” tasks
Characteristics for AGI Environments, Tasks, and Agents
- Complex dynamic open environment; diverse, interacting, richly structured objects
- Task-relevant regularities at multiple time scales
- Other agents impact performance
- Complex, diverse, novel tasks
- Interactions between agent, environment and tasks are complex and limited
- Computational resources of the agent are limited
- Agent existence is long-term and continual
Cognitive Architecture Requirements
New tasks do not require agent re-programming
•Realize a symbol system
•Represent and effectively use knowledge: modality-specific, diverse levels, large bodies, different levels of generality; beliefs independent of current perception; rich hierarchical control, meta-cognitive
•Support a spectrum of bounded and unbounded deliberation
•Support diverse, comprehensive learning and incremental, online learning
AGI: Full Development Theories
Vygotsky….
Piaget….
Substrata for AGI Landscape (Adams et. al, 2011)
Information processing - Implementation Metaphor
Math - Information Theory: formal definitions of intelligence, intelligence as compression
Physiology - Existing Realization: Insights from bioscience, universal aspects of development
Challenge or AGI Landscape
finding tasks and environments
Cognition implies…..
ability to understand how things might possibly be, now or at some future time
… using the past to predict the future (memory, anticipation)
… assimilating what does actually happen to adapt and improve anticipatory abilities
Cognition breaks free of the present
A cognitive system…
can function effectively under circumstances that were not planned for explicitly when the system was designed
Plasticity, resilience in the face of the unexpected
… can explain what it is doing and why it is doing it
Identification of potential problems, recognise need for new information
… can view a problem in more than one way and use knowledge about itself and the environment
Self-reflection
Traditional assumption of cognition
planning dominates over execution
Failing -> plan repair -> continuous planning and erring
Characteristics of traditional assumption
Boundedness, limited autonomy, multiple goals/tasks
–Bird-eye’s view first person point of view
–From full top-down control (“goal”, “plan”) towards perceptually coordinated activity ()
–Motivated configuration and interpretation of sensory-motor couplings to the world
–Distinction of subjective concerns from “goals”
•Persistence, social environment
–History matters! Adaptivity, individuality,… personae
New assumption
Offline certainties & reifications, closed worlds online qualities & dynamics, open worlds
Short summary of Cognitivist approach
Based on symbolic information processing representational systems
•“Cognition as type of computation” defined on symbolic representations
•Representations instantiated physically as cognitive codes
•Behaviour is a causal consequence of operations on these codes
•Cognition as symbolic, rational, encapsulated, structured, algorithmic
Short summary of Emergent System
- Emphasis on principles of self-organisation, emergence
- Dynamical systems view
- Connectionist systems, Enactive systems
Short summary of Hybrid Models
•Focus on perception-action behaviours rather than perceptual abstraction of representations
Modelling Approaches of Cognitivist, deliberative, reflective, symbolic
–Main aspects: inferencing mechanisms, representations, rationality, problem solving
–Cf. Folk Psychology/Theory Theory
–Well-known architectures:
•SOAR, EPIC, ACT-R, …
Grounding problem: Connecting to the “environment”
Modelling approaches of Situated, embodied, emergent, enactive
Main aspects: robust routine behaviour, understanding and exploitation of coupling to the environment
–Cf. John Dewey,…
Autonomy problem (separation from one’s environment)
Computational Operation
•Cognitivist:
–Rule-based manipulation (syntactic processing) of symbol tokens
–Typically sequentially
•Emergent:
–Concurrent interaction of a network of distributed interacting components
–Exploitation of self-organization, self-production, self-maintenance, self-development
Representational Framework
•Cognitivist:
–Patterns of symbol tokens that refer to events in the external world
–Typically descriptive product of human designer (perspective of observer)
–Usually punctate and local
•Emergent:
–Global system states, encoded in the dynamic organization of the distributed network of system components
Semantic Grounding
•Cognitivist:
–Symbolic representations grounded through percept-symbol identification by the designer or by learned association
–Representations are accessible to direct human interpretation
•Emergent:
–Representations grounded by autonomy-preserving anticipatory and adaptive skill construction
–Meaning only in terms of contributions to sustained system viability, not accessible to direct human interpretation
Temporal Constraints
•Cognitivist:
–Not necessarily entrained (=automatically synchronised) by events in the external world
•Emergent:
–Synchronous real-time entrainment with (events in) the environment
Inter-agent Epistemology
•Cognitivist:
–Positivist view of reality: absolute shared epistemology
–Structure and semantics of environment are independent of the system’s cognition
•Emergent:
–Agent-dependent epistemology as subjective outcome of a history of shared consensual experiences among phylogenetically compatible agents (species)
Embodiment
•Cognitivist:
–Not necessarily implied
–Cognition as distinct component of the brain (disembodied in principle)
•Emergent:
–Cognition implies embodiment
–Embodied perception and action (physical instantiation is constitutive)
Perception
•Cognitivist:
–Perception as interface between external world and internal symbolic representations
–Perception abstracts faithful spatio-temporal representations from sensory data
•Emergent:
–Change in system state in response to environmental perturbations in order to maintain stability
Action
•Cognitivist:
–Causal consequence of symbolic processing of internal representations
•Emergent:
–Perturbation of the environment by the system
Anticipation
•Cognitivist:
–Typically in the form of planning using procedural or probabilistic reasoning using a priori models
•Emergent:
–Self-effected traversal of self-constructed perception-action space
–Requires the system to visit a number of states without committing to the associated actions
Adaptation
•Cognitivist:
–Usually implies acquisition of new knowledge
•Emergent:
–Structural alteration or reorganisation to effect a new set of dynamics
Motivation
•Impinges on: –Perception (through attention) –Action (through action selection) –Adaptation (through factors governing change) •Cognitivist: –E.g.: resolving an impasse •Emergent: –E.g.: enlarging space of interaction
Relevance of Autonomy
•Cognitivist:
–Not necessarily implied
•Emergent:
–Crucial: Cognition as process whereby an autonomous system becomes viable and effective
what is Autonomy?
The ability of a system to contribute to its own persistence
•The degree to which a system’s behaviour is not determined by the environment
•The self-maintaining organizational characteristic of living creatures (dissipative far-from equilibrium systems) that enables them to use their own capacities to manage their interactions with the world, and with themselves, in order to remain viable
Cognitivist models
Cognitivism = information processing approach
–Cognition involves computations…
… defined over internal representations qua knowledge
–Information about the world is abstracted by perception…
… and represented in symbolic data structures Perceive Reason Plan Act (in the world)
–Ontological commitment to correspondence
–Perception abstracts faithful spatio-temporal representations
–Transduction problem: how can such reliable abstraction be achieved?
Representations
–Cognition is representational, it involves the manipulation of explicit symbolic representations of state and behaviour of the external world
–Reasoning is symbolic
–Source of symbolic representation (framework)s: human designers
–Directly accessible and understandable by humans
–Semantic knowledge can be entered and extracted directly
–Focus on goal-directed behaviour
Issues
–System blindness: The content of a cognitive system is constrained by (idealised) human perspective and human capabilities, the system “itself” is blind
–Brittleness of system: complete breakdowns upon encountering semantic gaps: what is not represented does not exist and therefore cannot be accessed
–Issues with execution (sensing and acting): inherently uncertain, time-varying, incomplete, …
PSS stands for
Physical Symbol System
Physical Symbol System Hypothesis
PSSs have the necessary and sufficient means for general intelligent action
Any system exhibiting general intelligence is a PSS
Any PSS of sufficient size can get to exhibit general intelligence
Heuristic Search Hypothesis
–Solutions to problems are represented as symbol structures
–Intelligence is exercised in problem solving by search
Emergent Models
Cognition: the process whereby an autonomous system becomes viable and effective in its environment
•Self-organization: continuous system reconstitution in real-time to maintain operational identity through moderation of mutual system-environment interaction and co-determination
•Co-determination:
–The cognitive agent is specified by its environment
–The cognitive process determines what is real/meaningful for the agent
The agent constructs its reality (world) as result of its operation in that world
Cognitive behaviour as automatic induction of an ontology inherently specific to the embodiment and dependent on the system’s experiences
Perception is concerned with the acquisition of sensory data in order to enable effective action
dependent on richness of action interface
–No isomorphic abstraction of the structure of an absolute external environment
Cognition and Perception in Emergent models
Cognition as complement of perception
–Perception deals with the immediate
–Cognition compensates for immediate nature of perception, dealing with longer timeframes: unfolds in real-time
Cognition is intrinsically linked with the ability to act prospectively: in the future, with what might be
Cognitive learning in emergent Models
–Primarily anticipative skill construction, not knowledge acquisition
–Required root capacity: processes guiding action and improving this capacity
What plays a pivotal role in emergent models?
Physical instantiation plays a pivotal role in cognition
Synchronous development in context of its environment places strong limitation on rate of ontogenetic learning
Emergent dynamical systems cannot be bootstrapped into advanced state of learned behaviour
Connectionist Emergent Models
- Parallel processing of non-symbolic distributed activation patterns
- Reliance on statistical properties rather than logical rules
- Dynamical systems which compute functions that best capture statistical regularities in training data
Dynamical Systems Emergent Models
Open dissipative non-linear non-equilibrium systems
•Dissipative: (cf. burning candle) diffusion of energy, phase space volume decreases over time
•Non-linear: dissipation is not uniform, only a small number of degrees of freedom contribute to the system’s behaviour: order parameters
•Non-equilibrium: cannot maintain structure or function without external sources: energy, material, information ( open)
•System: large number of related interacting components with large number of degrees of freedom
The behaviour of a high-dimensional system can be characterised with a low-dimensional model (order parameters)
Dynamical Systems Emergent Models (cont)
Typical properties achieved via self-organisation by dynamical laws
–Multistability, adaptability, pattern formation and recognition, intentionality, learning all without need for symbolic representations
–Examples of non-trivial behaviours: Perceptions of affordances, time-to-contact, figure-ground bistability, …
Generally: Not representation-hungry, but also no(t yet) higher-order cognitive faculties
Enactive Systems Emergent Models
“Nothing is pre-given” no need for symbolic representations
–Issues of importance for continued existence of the cognitive entity are brought out = enacted
–Co-determined by the entity interacting with the environment it is embedded in
Perception-action co-dependency
Enactive systems must be capable of effective action supporting continued system integrity
–The role of the cognition process is not: to abstract objective structure through perception and reasoning, but rather: to uncover or construct unspecified regularity and order that can be construed as meaningful because facilitating continued operation and development
–Enactive interpretation: Real-time context-based choice of relevance
The world experienced by the cognitive system is not independent of it!
First order systems
–Auto-poietic=self-producing: Physical identity through structural coupling
–Dependence on environmental perturbations for essential structural changes that permit to continue operating
Cellular systems, instinctive behaviour
Second order systems
–Operational closure: Structural coupling through a nervous system
–Enables association of internal states with different interactions of the organism
–Can perturb their own organizational processes and attendant structures
–Increased flexibility: Self-development in addition to self-production
Meta-cellular systems, ontogenetic behaviours
Third order systems
–Coupling between second-order systems
–Mutually shared structural adaptations give rise to new phenomenological domains: Language and shared epistemology
Social systems, communicative behaviours
Hybrid models
Focus on perception-action behaviours rather than perceptual abstraction of representations (e.g. for “animats”)
•Interaction as organising mechanism driving coherence of association between perception and action
•Representations to be constructed by the system itself
No meaningful direct access to internal semantic representations
Necessity of embodiment (at least for learning phase)
Issue: the “cognition-on-top-of-” syndrome, e.g.: “cognition-on-top-of-vision” (cf. transduction problem!)
