Task 3

Question 1

What is ACT-R?

Answer 1

= adaptive control of thought

a computer program that mimics human behaviour
-> is isolated from the external world

main components:
1. goal stack
2. current goal
3. procedural memory
4. declarative memory
5. outside world

Question 2

What is declarative memory?

Answer 2

• the things we know as facts
• a collection of chunks that contain a number of elements

Question 3

What is a chunk?

Answer 3

• declarative memory elements
• each main row of chunks holds a value in a different slot
• each chunk has a level of activation (= spreading information) -> not constant
  = a lot of activation - easy and quick to retrieve -> STM
  = low activation - hard to find -> LTM
  = no activation - effectively forgotten until activation level rises

Question 4

What is procedural memory?

Answer 4

• thinks we know how to do
• learned as production rules
  = different representations and goals
• stores procedures that have a condition (= if) and an action (= then) part

Question 5

What is retrieval request?

Answer 5

• connection from procedural to declarative memory
• production rule fired in procedural memory may require elements from the declarative memory

Question 6

What is production compilation? (+What are stages of learning production rules?)

Answer 6

• from declarative to procedural memory
• new production rules can be created in procedural rules from chunks in declarative memory
• stages of learning production rules
  
  1. understanding
  2. production compilation
  3. practice

Question 7

What is the goal stack ?

Answer 7

consists of goals that are not the immediate focus of attention, but that still need to be dealt with in the future
(= warteliste)

-> based on last-in-first out (LIFO)
-> goals can be pushed (added) or popped (removed)
-> can be criticised for its lack of psychological plausibility

Question 8

what is the current goal?

Answer 8

• focus of current attention
• representing what ACT-R is currently processing

Question 9

What are memory phenomena of ACT-R?

Answer 9

• list memory
  = experimental paradigm used to inderstand how things are stored in and recalled from STM
• forward recall
• backward recall
• free recall

Question 10

What is the primacy effect?

Answer 10

accuracy is the highest for the first elements in the list
-> participants rehearse the first elements of the list during the presentation of the other elements

Question 11

What is the recency effect ?

Answer 11

accuracy is the highest for the last elements in the list
-> last elements are still accessible from memory during recall phase (= activation level has not decayed)

Question 12

What does an ACT-R model have to do to accurately reflect empirical data?

Answer 12

1. Have a representation of how items are chunked in declarative memory
2. Have production rules for the rehearsal of items and retrieval from memory
3. Model activation levels and show how they affect recall accuracy and latency

Question 13

What are production rules ?

Answer 13

• condition part
  > retrieval from declarative memory (= function of activation level that matches the condition; the higher the easier)
• action part

-> similar to the action performed by a turing machine in a single step

Question 14

What is SNIF-ACT?

Answer 14

• scent based navigation and information foraging in the ACT architecture
• developed to stimulate users as they perform unfamiliar information-seeking tasks on WWW
• it selects actions based on the measure of information scent
  > calculated by a spreading activation mechanism that caputures mutual relevance of the contents of the webpage to the goal of the user

Question 15

what are the main predictions and results of SNIF-ACT?

Answer 15

1. link following behaviour:
   users working on unfamiliar tasks are expected to choose links that have high information scent
2. Points at which users will give up on WWW sites:
   when the information scent of the site diminishes below a certain threshold

results:
current content based spreading activation SNIF-ACT model can generate useful predictions about complex user-WWW interactions

Question 16

what is the association strength ?

Answer 16

• strength of the bond between an item and the required chunk
• influences the flow of activation between chunks
• depends on the total number of associations
• If an item is only associated with one chunk, this chunk receives the full associative strength of the item = the full effect of any activation

Question 17

What is the fan effect?

Answer 17

• applies to list memory
• the greater the number of facts related to some concept that a subject has to memorize, the slower the subject will be to recall any one of them.
• explained through limited capacity of association strength

Question 18

What is an the activation threshold?

Answer 18

• a chunk that falls below the activation threshold is unavailable for retrieval by the production rules
• The activation level affects recall success.
• ACT-R specifies the relationship between activation and latency = the weaker the activation, the slower the recall process
• partial matching is possible
  > important in positional confusion (= transportantion errors)

Question 19

What are auxiliary and architectural assumptions?

Answer 19

auxiliary: decisions on how to deal with flexibility
architectural: general claims to the nature of human cognition

advantage of ACT-R:
all assumptions are made explicit

Question 20

How does the user tracing method work?

Answer 20

• two tasks
  > Antz Task
  > City task (= easier)
• tracing instrumentation:
  > Weblogger (= tracking keystrokes, mouse-movements, button use, browser actions)
  > eye tracker
  > video-recordings of screen-display

Question 21

What are the results of the user tracing method?

Answer 21

• measure of information scent is able to generate good predictions to user-WWW interaction
• Most of the links chosen by the participants were ranked high by SNIF-ACT
  (ACT predict which links people will click on a web page)
• Information scent was shown to be able to predict when people will leave a site
  -> information scent of the site was decreasing

Question 22

How are lists declaratively represented?

Answer 22

• as chunks, as set of groups (groups of 3)
• each chunk is represented using slots/ values

Question 23

What is logic?

Answer 23

• the study of reasoning with definite knowledge
• important for our understanding of general purpose intellignce
• main requirement = formal language
• types
  > propositional logic
  > first order logic

Question 24

what is propositional logic?

Answer 24

• symbols (stand for propositions that can be true/ false)
• logical connectives (and, or, not, if…then)
• limitations
  > not very expressive
  > generalisation

Question 25

What is first-order logic?

Answer 25

• more expressive logical language
• assumes the world is made of objects that can be related to each other in various ways
• allows to assert that some property is true for all objects in the world
• first-order probabilistic language
  = ciombines first-order logic and probility

Question 26

what is probability?

Answer 26

• outcomes = possibilities wha t can occur after event
• probability of outcomes can change if they are dependent on previous outcomes

Question 27

what is a belief state?

Answer 27

• state in which we believe things to be
  > things can be uncertain (impossible to always know where everything is)
• need to be informed to make decisons
  > update information based on new input
• problem for AI:M hard to model not to know but to have high probability
• prediction state

Question 28

What is learning from examples?

Answer 28

• supervised learning (= deep learning)
  > most common form of machine learning
  > seeks to optimise the agreement between the hypothesis and the training examples
  > learning takes place by a sequence of modifications to the hypothesis

Question 29

What is a deep convolutional network ?

Answer 29

• Represents a complex mathematical expression composed in a regular way from many smaller subexpressions
• The compositional structure has the form of a network
• Convolutional as the network structure repeats itself in a fixed pattern across the whole input image
• “deep” because such networks typically have many layers

Question 30

What is CRUM?

Answer 30

• Thinking is best understood in terms of representational structures in mind and computational procedures that operate on them
• dominant approach in cognitive science
• Comparing mind with computers provides a powerful metaphor for mental processes
• Program (Data structures + algorithms) = running computer program
• Mind (mental representation + computer program) = thinking

Question 31

what is learning from thinking ?

Answer 31

• when first encountering a problem, you will think of a solution
• if you encounter a similar issue, you will store and reuse a generalised solution to the problem
• it is possible to learn effective, general rules from a single example
• in AI = explanation based learning

Question 32

which categories are important when Evaluating computational models?

Answer 32

Implement in computer programs and compare with human performance

a) Representational power
b) Computational power
c) Psychological Plausibility
d) Neurological plausibility
e) Practical applicability

Question 33

What is representational power?

Answer 33

• goals

Question 34

What is computational power?

Answer 34

• problem-solving
• planning
• decision making
• explanation
• learning
• language

Question 35

What is psychological plausibility

Answer 35

Rule-based systems have the most psychological applications of all computational-representational approaches

Question 36

What is neurological plausibility ?

Answer 36

• Anderson sketches a possible neural implementation of ACT
• simple rule-based systems have been implemented in artificial neural nets
• examples:
  > fMRI
  > EEg
  > MEG

Question 37

What is practical applicability? (ACT)

Answer 37

• Rule-based systems have been used to model learners’ performance
• used to build computer tutors that can help them learn
• example: snif ACT (based on ACTR)