Putting the Pieces Together
Example: A Complete Architecture:We have seen a number of different approaches to building a cognitive science
The complete science will be a complex structure that uses all or many of the parts
An evolutionary, modular approach
A rule based system using physical symbol systems
An associationist system using neural networks
ACT-R/PM developed by John Anderson:
HAM => ACT => ACT* => ACT-R => ACT-R/PM
Primary Modules:
Perceptual-motor (PM)
Memory: Declarative and Procedural
Buffers: representing current state of ACT-R
Pattern Matchers: relating productions to current buffers
ACT-R Components: Many elements we have already seen
See About ACT-R, the ACT web site
Modeling Performance
Dependent variables:Reaction times, accuracy, and fMRI data
An arithmetic task simplifying equations
Uses fMRI data to guess at subject's next response
Areas of Application
Creating a working model in ACT-RCognitive domains:
learning and memory
problem solving and decision making
language and communication
perception and attention
cognitive development
individual differences
Used in several areas of application, including Cognitive Tutors for mathematics: Widely used in schools
You can download this elementary tutorial to see more information on how ACT-R works. Note the following features:
Knowledge: Declarative knowledge is defined by"chunks"
Procedures: Procedural knowledge is defined by productions
Goals: Special importance is given to goals, which direct behavior
Knowledge acquisition: Productions can produce new chunks
Chapter 10: Section 10.1
What is an agent? What is an agent architecture? What are the distinguishing features of an agent architecture?
The goal of cognitive science is to describe mental architecture
Agents are the buildings of this architecture
Each agent will have, at a minimum, a sensory system and an effector (output) systemDescribe a simple reflex agent. What are production (condition-action) rules?
Not much more than the sensory and effector components
Productions are rules that generate an action for a given condition
Reflex: No real intelligence
But intelligence can emerge from higher level uses of these agentsWhat is a goal-based agent? How does it operate?
Goals are higher level motives than simple reflexes
Such agents understand consequences, and can compare consequences to goals
They can employ reflex agentsFrom Reflex Agents to Goal-Based Agents
A reflex agent is a simple production rule
How does ACT generate goal-based action out of production systems?
One of the ACT buffers is a goal buffer. Productions act on buffers
Use goals as conditions and actions
IF goal is add and 1, 1
THEN answer = 2 to output buffer
Create goal to output answerWhat is a learning agent?
Simpler goal based agents cannot learn from experience
Learning agents can improve their own performance
Note here how we have the outline of a system that can use both PSS and connectionist elementsLearning in Production Systems
How does a production system learn
Temporary changes through the buffers
Three elements can be changed permanently: Conditions,Actions, Priority ordering
How is the learning implemented
(a) Higher level productions?
(b) Have productions emerge as the output of neural networks more effective
Chapter 10: Section 10.2
What is a cognitive module? What are the characteristics of modular processing? Give some examples of candidates for cognitive modules.
Four important features (according to Fodor):
Specific to a single domain
Information is encapsulated independent of other activity
Mandatory application (when relevant)
FastTwo other features (less important, according to Fodor)
Sometimes associated with specific brain structure
Fails in a predictable wayAn interesting module: Face recognition
What is central processing? What does Fodor mean when he calls central processing Quinean and isotropic?
Fodor concludes there must also be domain-general, central processing
Produces knowledge and beliefs that characterize the organism as a whole
Quinean: Involves a theory of the world
Isotropic: The parts are inter-dependentWhere do Fodors cognitive modules tend to fall in agent architectures?
The modules operate at the interface of the cognitive system and the environment:
Perceptual processes and motor processes
For Fodor, modules are autonomous and independent of each other. They require the central processor to integrate everythingExplain Fodors First Law of the Non-existence of Cognitive Science. What is the tension between the language of thought hypothesis and central processing?
The more global a cognitive process, the less anyone understands it
Central processing requires that we solve the framing problem
And no-one knows how to do thatWhat are some ways of responding to Fodors argument for Fodors First Law of the Nonexistence of Cognitive Science?
The argument can be expressed as a seven step syllogism
As with any syllogism, the conclusion must be true if the premises are true
To falsify the conclusion we must falsify one or more of the premises1. The best model we have for understanding information processing is the LOT model
2. According to the model, information processing is manipulating sentences in LOT
3. Operations in LOT depend solely on syntax
4. Syntactic properties are physical properties
5. Central processing depends on global properties of a systems theory of the world
6. Global properties are not physical properties
7. Therefore, central processing cannot be understood within the LOT modelHow successful is Fodors argument for Fodors First Law of the Nonexistence of Cognitive Science?
Bermúdez takes aim at premises #5 and #6 by considering the massive modularity hypothesis
But what about premises #1 through #4?
Chapter 10: Section 10.3
What is the massive modularity hypothesis? What arguments have been offered to support the hypothesis?
Suppose there were no general purpose, central information processing ?
I.e., suppose everything is handled by special purpose modules
This was the proposal of Cosmides and Tooby
They suggest a swiss army knife analogy
Examples: Language; cheat detection
Swiss army knife: arguments from error and from statistics of learning
Only a modular system can profit from errors
Only a modular system can learn to identify relevant triggering stimuli
Kin recognition example:
Why would mother sacrifice herself for her children
She must be able to recognize (a) kinship and (b) relative benefits of her sacrificeTo what degree are Cosmides / Tooby modules innate?
It is often assumed that they come built in
Example: The language instinct
This is not necessary, however
Having some innate component makes it easier to explain a modules operation
Example: Language obviously involves learning, but very rapid, facilitated learningCompare and contrast Fodorian and Darwinian modules.
The Cosmides / Tooby modules are less obviously modular.
In terms of Fodors six criteria, only #1, specific to a single domain, clearly applies
Encapsulation, mandatory operation, speed may apply in some cases, but they are not absoluteExplain the domain-specific bodies of knowledge objection to the massive modularity hypothesis.
The Cosmides / Tooby modules are not modules in the sense used by Fodor
They are more like domain-specific bodies of knowledge
They may exist (and arguments for their existence are strong), but this does not rule out a general-purpose process as wellReconstruct the two arguments purporting to show that there must be domain-general processing?
Fodor: (a) How do modules identify their appropriate input? (b) How does the organism mediate between competing outputs?
These input/output concerns require a higher level, general purpose processing
(But do they?)Note similarity of this discussion to arguments over the nature of intelligence
General versus special purpose processing
An ingenious approach by Steven Mithen
Argues from paleontological evidence that hominids have gone through cycles of special purpose and general purpose skills
Example: Where are activities located in sites occupied by early humans?A related issue: The nature of self
If everything is processed by separate modules, where is you?
(This sort of question may motivate Fodors critique, and notion of theory of the world)
Daniel Dennett: This is a misleading question
You are the collection of processes
Conscious awareness is the end result of a battle among activities that compete for attentionHow good is Cosmides and Toobys case for the massive modularity hypothesis?
Another of the arguments that will never be settled?
Chapter 10: Section 10.4
Explain the basic setup and functioning of ACT-R/PM.
One way to settle many of the theoretical questions is to build a working system
ACT is modular, and combines physical symbol systems with non-symbolic implementationHow does ACT-R/PM represent declarative and procedural knowledge, respectively?
Declarative: Knowing that
Procedural: Knowing how
Declarative: associational structures in chunks
Procedural: production systems
Buffers connect cognitive elements with perceptual-motor elements
So far everything is symbolic and modularHow does ACT-R/PM select which production rule should be active if there is a conflict?
Pattern matching and rational choice
Each matching production rule has a utility
Benefit: likely goal achievement
Cost of activation
Hence the rational in ACT-RWhich parts of ACT-R/PM are symbolic? Which are subsymbolic? Explain why.
Utility calculations are one form of subsymbolic activity
They do not reflect the symbolic content of the productions
Activation of chunks in declarative memory is also subsymbolic
Chunks are symbolic. Activation levels are not