Course Overview

1
Course Overview

• What is AI?
• What are the Major Challenges?
• What are the Main Techniques?
• Where are we failing, and why?
• Step back and look at the Science
• Step back and look at the History of AI
• What are the Major Schools of Thought?
• What of the Future?

?Done
Part IIGive you an appreciation for the big
picture ? Why it is a grand challenge
2
Course Overview

• What is AI?
• What are the Major Challenges?
• What are the Main Techniques?
• Where are we failing, and why?
• Step back and look at the Science
• Step back and look at the History of AI
• What are the Major Schools of Thought?
• What of the Future?

?Done
Part IIGive you an appreciation for the big
picture ? Why it is a grand challenge
3
Course Overview

• What is AI?
• What are the Major Challenges?
• What are the Main Techniques?
• Where are we failing, and why?
• Step back and look at the Science
• Step back and look at the History of AI
• What are the Major Schools of Thought?
• What of the Future?

• Looking at the Science
• Engineering vs. Science
• Introduction to Cognitive Science
• Cognitive Psychology 1
• Cognitive Psychology 2
• Cognitive Development
• Linguistics
• Neuroscience
• Philosophy

4
Course Overview

• What is AI?
• What are the Major Challenges?
• What are the Main Techniques?
• Where are we failing, and why?
• Step back and look at the Science
• Step back and look at the History of AI
• What are the Major Schools of Thought?
• What of the Future?

• Looking at the Science
• Engineering vs. Science
• Introduction to Cognitive Science
• Cognitive Psychology 1
• Cognitive Psychology 2
• Cognitive Development
• Linguistics
• Neuroscience
• Philosophy

5
How are Concepts Defined?

• Classical Theory define necessary and sufficient
  conditions
• Grandmother a female who has a child who has a
  child
• Likely properties are neglected grey hair, old
• Difficulties
• Not so realistic perhaps
• We dont have a clear idea of conditions for most
  concepts
• Old woman with adopted son who has children
• Usually consider her a grandmother
• Introspection suggests
• We often classify using unnecessary features
• Dogs 4 legs and barks
• Even though a dog who has lost a leg, and lost
  his voice, is still a dog!
• To be sure of dog we should
• have a careful assessment of its morphology, or
  chromosomes
• but this is not how we work

6
How are Concepts Defined?

• Deficiencies in classical theory
• Prototype approach
• Cognitive Scientists move to a likelihood
  theory
• Likelihood that a concept will have some
  characteristics
• Likelihood that something is categorised as that
  concept
• Members of a concept have Family Resemblance
• Family Resemblance idea picks very typical
  features
• Bird
• Robin has very typical features
• Flight, size, tendency to perch in branches, sing
• Penguin does not
• Methods to implement
• Likelihood schema
• Set up a schema with likely features,
• and weights on importance
• Or use Average of known examples

7
Evidence for Prototype Concepts

• Experiment by Rosch and Mervis
• Took categories
• Fruit
• Vegetables
• Clothing
• Furniture
• Vehicles
• Weapons
• Subjects given 20 items that were instances of a
  category
• Asked to list typical features
• From subjects responses each item was given a
  family resemblance score
• For each item One point for each feature also in
  another item
• E.g. furniture chair scored highest, telephone
  lowest

8
Evidence for Prototype Concepts

• Another experiment
• Subjects given typical and atypical instances of
  a category
• e.g. furniture chair, rug, table, telephone
• Asked to rate them on a 1-7 typicality scale
• Items with highest family resemblance score
  (from previous) given highest rating
• Shows
• having features in common with other members
  means more typical

9
Evidence for Prototype Concepts

• Another experiment
• Subjects given a category, and then instances
• Asked if instance belongs yes or no
• e.g.
• Bird robin ? yes
• Bird rabbit ? no
• Items with highest family resemblance score had
  faster response
• e.g.
• Bird robin ? fast
• Bird pigeon ? medium
• Bird eagle ? medium
• Bird chicken ? slow
• Shows
• having features in common with other members
  means more typical

10
How do we Choose Concepts

• Rosch analysed features we use
• Typical use
• Visual shape
• Suggests these characteristics constrain
  categories
• Culture (use)
• Visual system

11
How to Represent Concepts

• Can use propositions as before
• Proposition represents both the item and the
  concept
• Example
• Vegetable plant plant green
  bean edible edible fibrous fibrous
  green green main dish main dish long/thin
• Put a weight on each link
• to indicate how important it is to distinguish
  that concept
• Check
• how many overlapped paths
• And how strong
• To decide in green bean is a vegetable

12
Dynamic Theory of Concepts

• Proposed by Barsalou 1993
• When concepts retrieved in a certain context
• Certain features are given prominence
• Example thinking of concept cucumber
• During Spring planting
• During August dinner
• Different features given prominence
• Experimental evidence
• Subjects were given a context with a sentence
  (priming)
• Then asked if a feature was part of the concept
• Results showed low-weight features could be
  boosted
• Dynamic concepts
• Means that your notion of the concept is changing
• Depends on your current context

13
How to Learn the Concepts

• For a prototype concept
• Train a network with the examples that have been
  seen
• Adjust the weights on the features on the concept
• End up with a good average prototype
• Problem
• What about features like colour of a cow?
• Seem to be set of possible colours
• Not just any colour, but certain options

14
How to Learn the Concepts

• Exemplar Approach
• Alternative to prototype approach
• Store all the examples
• e.g. all known example of dog
• When a new one comes along, see how well it
  matches known ones
• dog-similarity value
• Approach works well in lab tests
• Better than prototype approach
• Concern
• Need to store so many examples,
• and compare a new instance with each stored one
• Could compare in parallel by neural network
• but still a lot of storage

15
How to Learn the Concepts

• Top down and bottom up processes
• Seeing a fat man in a foreign country
• You would not conclude that all men in that
  country are fat
• Seeing a coin in a foreign country
• You would conclude that all those coins would
  have that size
• This is using some higher level knowledge
• People seem to have theories of domains
• Concepts seem to incorporate high level
  knowledgeas well as low level likely features
• Proper theory of concepts may take some time

16
Cognitive Science Concepts and AI?

• Sometimes the devil is in the details
• It is easy to describe for some simple concepts
  and features
• Describe a handful, and how they link in an
  associative network
• but does not scale up for a great number of
  concepts
• Number of features seems infeasible
• Example Barsalou has can be walked on as a
  feature of roof
• Imagine how many features roof has if we want to
  go to this level of explanation
• When it comes to connecting to the world
• Not clear how to do it
• Even recognising the most basic things is beyond
  vision systems
• A chair
• Unless constrained to particular types/lighting
  etc.
• Recognising most basic concepts from language
  also problematic
• Concepts most interesting in toy demonstrations
• Conclusion Cognitive Science Concepts
  interesting
• Clearly reveals some insight on how mind works
• But still a big gap between them and AI systems

17
Memory

• We will focus on declarative memory
• i.e. think of declare some fact to be true
• We already talked about procedural memory
• Skill acquisition play musical instrument, ride
  a bicycle
• Psychologists consider three stages for memory
• Acquisition
• Retention interval
• Seconds, minutes, years
• Retrieval
• Short-term / long-term
• Think of difference between your own phone number
• And one you remember just long enough to dial
• Experiment
• Subjects asked to try rehearsal or
  elaboration
• Rehearsal was good for short-term recall
• Elaboration was good for long-term recall
• Why?

18
Memory

• If subjects do deeper processing
• have better long-term memory recall
• Experiment
• Is the word in capital letters? table TABLE
• Does the word rhyme with weight? crate MARKET
• Is the word a type of fish? SHARK heaven
• Would the word fit this sentence FRIEND cloudHe
  met a _____ in the street
• Subjects answered 40 questions on different words
• Result words where the question required deeper
  processing were remembered better
• Also experimented with higher complexity
  sentence questions
• Even better memory
• Interesting intention to remember does not help!
• Another experiment
• Some subjects told they need to remember
• Others told they just need to answer quickly,
  then given surprise memory test at end

19
Memory - Elaborations

• In terms of propositional associative networks
• Elaboration activates more connected nodes
• If you forget the main part, the associated
  activations might activate it
• Some elaborations produce better memory effects
  than others
• Bradshaw and Anderson showed cause and effect
  effective
• Mozart made a long journey from Munich to Paris
• Cause Mozart wanted to leave Munich to avoid a
  romantic entanglement
• Effect Mozart was inspired by Parisian musical
  life
• Downside of elaborations
• Subjects often remember things that werent there
• After 24 hours
• Subjects recalled 1 incorrect elaboration for
  every 2 propositions in the story
• Relevant to witness testimony
• Watergate John Dean misattributed statements to
  people
• Subjects shown film of car crash
• Asked how fast were they going when they
  smashed into each other? how fast were they
  going when they hit each other?
• First group more likely to have seen broken
  glass

20
Reasoning

• Remember deduction from the AI part on logic?
• IF a guy is tall THEN Mary likes the guy
• John is a tall guy-------------------------------
  ----------------------
• Mary must like John
• Do humans really use logical deduction?
• Experiment Four cards
• E K 4 7
• IF a vowel on one side THEN must be an even
  number on other side
• High rate of error
• But performed better if detecting cheating
  involved
• Deductive model should not depend on content
• Why are humans so bad at logical reasoning?
• Human thought more heuristic works most of the
  time

21
Course Overview

• What is AI?
• What are the Major Challenges?
• What are the Main Techniques?
• Where are we failing, and why?
• Step back and look at the Science
• Step back and look at the History of AI
• What are the Major Schools of Thought?
• What of the Future?

• Looking at the Science
• Engineering vs. Science
• Introduction to Cognitive Science
• Cognitive Psychology 1
• Cognitive Psychology 2
• Cognitive Development
• Linguistics
• Neuroscience
• Philosophy