Task 5 Flashcards
What is ACT?
It is a cognitive model based on human associative memory theory- it examines how memories are represented and how those representations mediate behavior.
Problem: ACT only takes declarative and not procedural memory into account
Solution: ACT-R was proposed
What is ACT-R?
it is based on the rational analysis theory - each component of the cognitive system is optimized with respect to demands from the environment, given its computational limitations
Components of ACT-R: procedural memory
PROCEDURAL KNOWLEDGE/MEMORY (skills): PRODUCTION RULES are used to represent it (if…then).
- for a rule to be used, the IF part must match the current goal & use chunks already available in declarative memory
- each production = condition-action pair
a) Condition (if): match current goal & use available chunks
- have activation too: if its high –> often used; if its low –> not used, even if it would match the foal state
b) Action: things to do when condition applies –> action can change goal state
Components fo ACT-R: declarative knowledge
DECLARATIVE KNOWLEDGE (explainable): CHUNKS are used to represent it
- schema like structures consisting of isa pointer specifying their category and some numbers of additional pointers encoding their content
- chunks are linked to each other through their value in a network (influence each other through spreading activation).
Associative strength
the strength of bonds between the item and required chunk influences the flow of activation between the chunk
Activation ACT-R
Activation is a limited source: if 1 item is associated with only one chunk, then this chunk receives full associative strength
Activation level: sum of chunks base-level activation + associative activation. - easily retrieved if a chunk is highly active (STM)
- harder to retrieve when it has a lower activation level (LTM)
In order for a chunk to be retrieved it has to reach an activation threshold
Partially matching
When a partially matching chunk (it does only match somewhat to the current goal state) has a high activation level and a fully matching chunk is absent or below the activation threshold
- can explain positional confusion: one recalls a correct item but in the wrong position
Base-level activation (chunks)
Depends on the number of times a chunk has been rehearsed + the amount of time that has elapsed since it was last rehearsed
- accounts for primacy and recency effect
Information flow ion ACT-R
Goal stack: encodes hierarchy of intentions guiding behavior
Current goal: presents focus of attention at the moment (at any time, only 1 current goal). It can be pushed or popped).
- computer stacks can be used to hold a large number of elements + recall them perfectly (lacks psychological plausibility)
Conflict resolution: productions are selected to fire through a conflict resolution process -> choose one production among many that fit goal.
- created in procedural memory from chunks coming from declarative memory
- selected production can cause action in the world, transform the current goal or make retrieval requests from declarative memory
Chunks enter declarative memory either via:
a) popped goals: reflect the solutions from past problems
b) perception from the environment
Product compilation: Products are created from declarative chunks.
Process by which production rules are learned in ACT-R:
1. understanding: of instructions and examples
2. product compilation: apply instructions through generalizing from experience
3. Practice
Origin of chunks
Chunks originate from
a) production rules: which in turn originate from chunks
b) encoding from environment: ACT-R is a fundamentally sensationalist theory in that chunks result from environmental encoding:
1. ACT-R can move its attention over the visual array of objects
2. Once synthesized, such objects are available as chunks in WM of ACT-R
3. calls for shifts in attention result from firing of production rules
Basic assumption: process of categorizing an object from a set of features is equal to process of categorizing a visual pattern from a set of features
Example: chunk encoding letter ‘H’ will recognize its parts first while not recognizing the whole
Origin of production rules
mimic examples of solutions (transformations in the environment)
How does ACT-R selects the appropriate knowledge in the right context?
KNOWLEDGE DEPLOYMENT: increase in knowledge = slower processing
Initial parallel activation process: identifies the knowledge structures that are most useful in the context (chunks & rules)
- implicitly performs bayesian inference in calculating the offs of knowledge being used in a particular context
- for that it looks at the ACTIVATION LEVEL of chunks
Examples of knowledge deployment of ACT-R
- Memory
MEMORY: history and context combine to determine the relevance of a certain info
List memory: experimental paradigm used to investigate how people store & recall info
a) Forward recall: recall items in precise order
b) backward recall: recall reverse order
c) Free recall: any order
Group: the list itself is represented in declarative memory as chunks/groups
- a list is remembered by making smaller groups of the numbers (if you have 9 numbers and 3 groups, you need 12 chunks to remember it)
Slots and values: each chunk is represented by slot for tis group, its position within the group and the overall list
Goal transformation: transform goals with help of production rules
a) modify b) create c) remove
Generalization: variables in production rules act as empty slots & accept many specific values
Examples of knowledge deployment of ACT-R
-Categorization
When people are asked to classify patients as having a rare or common disease, they adhere to rates of disease and likelihood ratio of symptoms (same as activation level of chunks)
Limitation of ACT-R
Architectural constraints: limit how it works
- procedural declarative distinction
- use of chunks to group items in declarative memory
- auxiliary assumptions: decisions on how to deal with leftover flexibility
Cognitive modelling
Deals with simulating problem solving and mental processing in a computerized model. Models are used to predict human behavior
- serve to bridge behavior & neural underpinnings
Pros & Cons of cognitive modelling
PROS:
- produce logically valid predictions
- make quantitatively precise predictions
- Generalizability: make predictions that go beyond the original data
CONS: Not as neurobiologically precise as neural models
Approaches of cognitive modeling
- high-level
- HIGH-LEVEL (RULE-BASED SYSTEM): something should look & behave as much as a car as possible whiteout necessarily having the same initial workings
Cognition is modeled with an explicit set of rules: Production rules that contain
a) condition
b) action
WHY use rule-based models?
Representational power: the structure is simple, but can represent many different kinds of knowledge
Computational power: problem solving, learning and language
Approaches of cognitive modeling
- low level
- LOW-LEVEL (CONNECTIONISM): representing the kinds of bits cars are made of and try to understand how these components work together to behave like a car
Parallel distributed processing: match human performance by programming artificial neurons into networks
Steps of cognitive modeling:
- reformulate a theoretical framework into computer language
- make additional detailed assumptions if theory is too weak to completely specify a model
- Estimate parameters that are initially unknown
- Compare predictions of competing models
- all of them are wrong to some extent as they only want to capture one aspect . Try to find the model that better represents the system we are trying to represent - Reformulate theoretical framework and start constructing new models
Presentation article: Rules
- Representational power
- Computational power
Representational power: how much knowledge can a rule-based model represent
Computational power: how powerful and efficient are rule-based models
Cognitive constraints of decision making under uncertainty:
Bounded rationality approach
Bounded rationality approach: optimal decision making is bounded by limitations of decision maker (these constraints should be modeled In a computational framework)
AIM: highlight importance of applying the same modeling approach to widely different paradigms of decision making
MODEL: Anticipate next stimulus location & time leads to shorter reaction time. Chunks of small pieces of stimulus sequence are represented in WM. Ability to learn sequences depends on the length of interval btw a response and the next stimulus
APPLIED TO PAPER-ROCK-SCISSOR: humans find it harder to generate random actions but easy to detect sequence frequencies. (best player would make random actions).
CONCLUSION: application of the same model across many paradigm illustrates predictive benefits of models based on Cognitive architectures
