Cognitive Architectures Flashcards
- How to implement a cognitive system? - Cognitive Architecture – A Definition - Classification of Cognitive Architectures - Requirements Analysis - ACT-R – A Cognitive Architecture
What is a cognitive architecture?
A cognitive architecture is a model of a cognitive system that defines the basic and irreducible cognitive operations that enable the human mind.
(In theory, each task that humans can perform should consist of a series of these discrete operations.)
(A cognitive architecture is a theory about the structure of the human mind and to a computational instantiation of such a theory.)
How was the design of cognitive architectures motivated?
The design of cognitive architectures was originally motivated in cognitivist cognitive science by Newell’s concept of a unified theory of cognition.
What is a cognitive model?
A cognitive model is an instantiation of a cognitive architecture with a specific body of knowledge.
What are the goals of cognitive architectures?
Modelling of the human mind
- advance understanding of cognition through computer-executable description of cognitive processes
- reproduction of experimental data
- focus on generality and robustness
Construction of artificial cognitive systems
- design of artificial systems that are as capable as living creatures
What are the advantages of cognitive architectures?
Cognitive architectures are…
- generic and can be applied to different domains
- unified models that coherently integrate many different aspects of cognition
- executable implementations of cognition theories
What are the levels of modeling regarding cognitive systems?
- sociological level: interaction between agents
- psychological level: cognition in individual agents
- componential level: cognitive architecture
What do cognitivist cognitive architectures model?
A cognitivist cognitive architecture models those aspects of cognition that are task-independent and constant over time.
- specifies only basic components and processes such as memory, knowledge
- generic model that reflects a specific set of assumptions
- the architecture is a framework that does not include any domain knowledge required for cognitive processing
What are production systems in cognitivist cognitive architectures?
A production system is a set of rules of the form if-then
- conditions are propositions on system components
- production systems are Turing complete and can adapt
- a production is activated (its action executed) when all of its conditions are satisfied
- when more than one production rules are satisfied, precedence rules are applied
What are emergent cognitive architectures?
Emergent cognitive architectures are frameworks that facilitate development. They model the innate skills of a newborn agent (i.e. its phylogenetic configuration).
- body is active part of the architecture
- during ontogeny (development), the architecture is instantiated through the acquisition of experience
- the architecture needs to provide a structure that accommodates different cognitive functions
- as a result of development, the architecture can change over time
What are design considerations for cognitive architectures?
Scope (compactness and comprehensiveness)
- cognitive architectures are generic and minimalistic -> not desirable to model all relevant cognitive functions
Generality (simplicity and realism)
- generality is an important tradeoff: too general means that the architecture will be under-constrained and meaningless
What are requirements for cognitive architectures?
Realism
- ecological realism: capability of an agent to engage in everyday activities of its natural environment
- bio-evolutionary realism: model of human intelligence should be reducible to a model of animal intelligence
- cognitive realism: capture of all essential characteristics of human cognition
- eclecticism: incorporation of results from prior research
Behavioral characteristics
- reactivity: quick reactions through immediate and fixed responses without elaborate cognitive processing
- sequentiality: sequential execution of activities
- routineness: learning, adaptation and execution of recurring patterns e.g. by trial and error
Cognitive characteristics:
- explicit processes: accessible and precise
- implicit processes: inaccessible and imprecise
- synergistic interaction: implicit & explicit processes interact to complement and supplement each other
- learning: implicit and explicit processes interact during top-down learning (explicit learning first, implicit learning later) and bottom-up learning (implicit learning first, explicit learning later)
- modularity: some cognitive faculties are specialized and separated through functional or physical encapsulation
Development
- value systems and motives: guidance during action selection and motivational drive for development
- physical embodiment
- sensorimotor contingencies: store relations between actions and resulting sensory feedback
- perception: capacity for unsupervised perceptual categorization
- constitutive autonomy: autonomous operation of developmental processes
What is the requirement of ecological realism?
Requirement for cognitive architectures / Realism
capability of an agent to engage in everyday activities of its natural environment
What is the requirement of bio-evolutionary realism?
Requirement for cognitive architectures / Realism
the model of human intelligence should be reducible to a model of animal intelligence
What is the requirement of cognitive realism?
Requirement for cognitive architectures / Realism
capture of all essential characteristics of human cognition
What is the requirement of eclecticism?
Requirement for cognitive architectures / Realism
incorporation of results from prior research
What is the requirement of reactivity?
Requirement for cognitive architectures / Behavioral characteristics
quick reactions through immediate and fixed responses without elaborate cognitive processing
What is the requirement of sequentiality?
Requirement for cognitive architectures / Behavioral characteristics
sequential execution of activities
What is the requirement of routineness?
Requirement for cognitive architectures / Behavioral characteristics
learning, adaptation and execution of recurring patterns e.g. by trial and error
What are explicit and implicit processes?
Requirement for cognitive architectures / Cognitive characteristics
- explicit processes: accessible and precise
- implicit processes: inaccessible and imprecise
What is the requirement of synergistic interaction?
Requirement for cognitive architectures / Cognitive characteristics
implicit and explicit processes interact to complement and supplement each other
What is top-down and bottom-up learning?
Requirement for cognitive architectures / Cognitive characteristics
- top-down learning: explicit learning first, implicit learning
- bottom-up learning: implicit learning first, explicit learning later
What is the requirement of modularity?
Requirement for cognitive architectures / Cognitive characteristics
some cognitive faculties are specialized and separated
What is the requirement of physical embodiment?
Requirement for cognitive architectures / Development
the body enables actions and to perceive the effects of actions
What is the requirement of sensorimotor contigencies?
Requirement for cognitive architectures / Development
sensorimotor contingencies store the relations between actions and resulting sensory feedback
What is the requirement of perception?
Requirement for cognitive architectures / Development
capacity for unsupervised perceptual categorization
What is the requirement of constitutive autonomy?
Requirement for cognitive architectures / Development
autonomous operation of developmental processes
What is ACT-R?
Adaptive control of thought – rational is a hybrid cognitive architecture implementing the principles of a unified theory of cognition.
ACT-R is based on modules, buffers, and a production system.
The theory is implemented as a Common Lisp framework.
What are ACT-R modules?
Modules in ACT-R are independent parallel processing units that encapsulate specific cognitive functions.
Modules are independent from another. Every module can contain arbitrarily complex subfunctions.
Four Main modules:
- intention module: maintains internal context of intentions for goal-directed behavior; current goal is exposed via the module buffer
- declarative module: declarative memory
- visual module: visual perception
- manual module: implements simple hand model with virtual keyboard/mouse
In addition:
- procedural module: implements the procedural memory system that manages, matches and executes productions. Coordinates interactions between all other modules
What are ACT-R buffers?
Buffers represent the current state of a module.
Every buffer stores one chunk and is assigned to exactly one module. It is visible to all other modules.
Buffers directly process queries about their content:
- empty
- full
- failure (true iff empty & failure flag is set)
- requested (true iff chunk in the buffer is the result of a request)
- unrequested (true iff chunk in the buffer is not the result of a request)
What are ACT-R chunks?
ACT-R chunks represent declarative knowledge. They are composed of slots, each of which stores a single value (which can also be another chunk).
Example: 8 + 4 = 12
add84
- addend1 Eight
- addend2 Four
- sum Twelve
(diagram: add84 ellipse, Eight/Four/Twelve squares, arrow add84 –sum–> Twelve)
Example: “The black cat with 5 legs sits on the mat”
cat007
…
fact007
…
fact007 –agent—> cat007
What is the ACT-R production system?
The ACT-R production system coordinates each module through its buffers.
Rules in the production system correspond to procedural knowledge
The production system operates in a critical cycle with a length of 50 ms
- pattern matching: matching of buffer values with production conditions
- selection and execution of a matching production
- results of the production are written back to the corresponding module buffers
What are ACT-R productions?
Productions are statements about behavior. They can be represented as if-then rules.
Productions are stored in and managed by the procedural module which monitors the state of all module buffers and automatically selects and executes a matching production
Production syntax: every production is identified by a name and is assembled from a list of buffer operations.
(p
==>
)
What is the condition of an ACT-R production comprised of?
The condition of a production is comprised of a set of buffer tests that are
- patterns which are matched against the current contents of the buffers
=buffername>
ISA type
slot1 const1
slot2 =var1
- queries for a buffer and module state information
?buffername>
buffer
What is the action of a production comprised of?
The action of a production is comprised of a set of
- buffer changes
=buffername>
ISA type
slot1 const1
slot2 =var1
- module requests
\+buffername>
ISA type
slot1 const1
slot2 =var1
What is the declarative module and how does it work?
The declarative module contains the declarative memory in the form of chunks.
- requests: search internal memory and return chunk with best match
- new memories are collected automatically from the buffer when a module is cleared
- two modes of memory retrieval: purely symbolic matching, time-aware matching
init syntax:
(add-dm
(b ISA …)
)
How is a chunk activation calculated?
Every chunk i in the declarative memory is assigned an activation Aᵢ:
Aᵢ = Bᵢ + Sᵢ + Pᵢ + 𝜀ᵢ
Bᵢ: base-level activation: recency and use frequency of the chunk
Sᵢ: spreading activation: effect on the contents of the module buffers
Pᵢ: partial matching value: matching-degree of the chunk for the request
𝜀ᵢ: noise value with a transient and permanent component
How is the base-level activation calculated?
Bᵢ: base-level activation: recency and use frequency of the chunk
Bᵢ = ln (∑ⱼ₌₁ⁿ tⱼ⁻ᵈ) + βᵢ
n: number of presentations of chunk i
tⱼ: time since the j-th presentation
d: decay parameter
βᵢ: constant offset
How is the spreading activation calculated?
Sᵢ: spreading activation: effect on the contents of the module buffers
Sᵢ = ∑ₖ ∑ⱼ Wₖⱼ Sⱼᵢ
Wₖⱼ: amount of activation from source j in buffer k
Sⱼᵢ: association strength from source j to chunk i
How is the partial matching value calculated?
Pᵢ: partial matching value: matching-degree of the chunk for the request
Pᵢ= ∑ₖ Pᵢ Mₖᵢ
Pᵢ: match scale parameter that determines the importance of similarity
Mₖᵢ: similarity value between the desired and the actual slot value
What is the procedural module and how does it work?
The procedural module implements the procedural memory system that manages, matches and executes productions. It coordinates interactions between all other modules.
- at every time step, the system computes the set of viable production rules by matching the buffer tests against the current buffer contents
- only one production rule can fire in every time step
Every production has a sub-symbolic utility value Uᵢ
How is the sub-symbolic utility value U_i calculated?
Every production has a sub-symbolic utility value Uᵢ that can be both set manually or learned online during operation:
Uᵢ = Uᵢ(n−1) + α[Rᵢ(n) − Uᵢ(n−1)]
α: learning rate
Rᵢ(n): reward for production i for its n-th usage
What is production compilation?
The production compilation mechanism tries to combine productions that fire in sequence.
What advantages does ACT-R have in contrast to a normal programming language?
- it advances the understanding of human cognition
- it is very generic and can be applied to different domains
- deep explanations of cognitive processing go beyond the task level
What are the two types of knowledge and their corresponding representation in ACT-R?
- declarative knowledge - represented by chunks
- procedural knowledge - rules of the production system
What is the difference between implicit and explicit learning?
When the process used to gain information is known, it is called explicit learning. It differs from implicit learning because of the awareness factor.
ACT-R model for counting
Task: Count from num1 to num2
Chunk Types (chunk-type count-order first second) (chunk-type count-from start end count)
Initialize Declarative Memory
(add-dm
(b ISA count-order first 1 second 2)
(c ISA count-order first 2 second 3)
(d ISA count-order first 3 second 4)
(e ISA count-order first 4 second 5)
(f ISA count-order first 5 second 6)
(first-goal ISA count-from start 2 end 4)
)
Set the Initial Goal
(goal-focus first-goal)
Define the Start Production
(p start
=goal>
ISA count-from
start =num1
count nil
==>
=goal>
ISA count-from
count =num1
\+retrieval>
ISA count-order
first =num1
)
Define the Increment Production
(p increment
=goal>
ISA count-from
count =num1
- end =num1
=retrieval>
ISA count-order
first =num1
second =num2
==>
=goal>
ISA count-from
count =num2
\+retrieval>
ISA count-order
first =num2
)
The Stop Production
(p stop
=goal>
ISA count-from
count =num
end =num
==>
-goal>
)
