animated version
Download
Report
Transcript animated version
Course Overview
What is AI?
What are the Major Challenges?
What are the Main Techniques?
Part I:
Introduce you to
what’s happening in
Artificial Intelligence
Where are we failing, and why?
Done
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?
Part II:
Give you an
appreciation for
the big picture
Why it is a
grand challenge
Course Overview
What is AI?
What are the Major Challenges?
What are the Main Techniques?
Part I:
Introduce you to
what’s happening in
Artificial Intelligence
Where are we failing, and why?
Done
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?
Part II:
Give you an
appreciation for
the big picture
Why it is a
grand challenge
The Language of Thought
What is the language we “think in”?
Is it our natural language, e.g. English, or mentalese?
Some introspective arguments against natural language
Word is “on the tip of my tongue”, but can’t find it
Difficult to define concepts in natural language, e.g. dog, anger
We have a feeling of knowing something, but hard to translate to
language
Some observable evidence against natural language
Children reason with concepts before they can speak
We often remember gist of what is said, not exact words
Cognitive science experiment: (recall after 20 second delay)
He sent a letter about it to Galileo, the great Italian Scientist.
He sent Galileo, the great Italian Scientist, a letter about it.
A letter about it was sent to Galileo, the great Italian Scientist.
Galileo, the great Italian Scientist, sent him a letter about it.
Represent as Propositions
Just like the logic we had for AI
isa
likes(john,mary)
likes
a
apple
gives
mary
john
john
a
mary
Evidence for Propositions
A cognitive Science experiment (Kintsch and Glass)
Consider two different sentences,
but both with three “content words”
The settler built the cabin by hand.
The crowded passengers squirmed uncomfortably.
Evidence for Propositions
A cognitive Science experiment (Kintsch and Glass)
Consider two different sentences,
but both with three “content words”
The settler built the cabin by hand.
One 3-place relation
The crowded passengers squirmed uncomfortably.
Three 1-place relations
Subjects recalled first sentence better
Suggests it was simpler in the representation
(Cognitive Science involves a fair bit of guessing!)
Associative Networks
Idea: put together the bits of the propositions that are similar
Associative Networks
Idea: put together the bits of the propositions that are similar
isa
likes
gives
a
mary
apple
john
john
a
mary
Associative Networks
Idea: put together the bits of the propositions that are similar
likes
isa
mary
john
a
apple
gives
Associative Networks
Idea: put together the bits of the propositions that are similar
Each node has some level of activation
Activation spreads in parallel to connecting nodes
Activation fades rapidly with time
A node’s total activation is divided among its links
Associative Networks
Idea: put together the bits of the propositions that are similar
Each node has some level of activation
Activation spreads in parallel to connecting nodes
Activation fades rapidly with time
A node’s total activation is divided among its links
These rules make sure it doesn’t spread everywhere
Nodes and links can have different capacities
Important ones are activated very often
Have higher capacity
These ideas seem to match our intuition from introspection
One thought links to another connected one
Associative Networks
Cognitive Science experiment (McKoon and Ratcliff)
Made short paragraphs of connected propositions
Subjects viewed 2 paragraphs for a short time
Subjects were shown 36 test words in sequence
and asked if those words occurred in one of the stories
For some of the 36 words, they were preceded by a word from
same story
For some of the 36 words, they were preceded by a word from
other story
Word from same story helped them remember
…Suggests it is because they were linked in a network
They also showed recall was better if closer in the network
…Suggests activation weakens as it spreads
Schemas
Propositional networks can represent specific knowledge
John gave the apple to Mary
…but what about general knowledge, or commonsense?
Apple is edible fruit
Grows on a tree
Roundish shape
Often red when ripe…
Could augment our proposition network
Add more propositions to the node for apple
Apple then becomes a concept
The connections to apple are a schema for the concept
What about more advanced concepts/schemas like a trip to a
restaurant?...
Scripts
Elements of a script…
Identifying name or theme
Eating in a restaurant
Visiting the doctor
Typical roles
Customer
Waiter
Cook
Entry conditions
Customer hungry, has money
Scripts
Sequence of goal directed scenes
Enter
Get a table
Order
Eat
Pay bill
Leave
Sequence of actions within scene
Get menu
Read menu
Decide order
Give order to waiter
Scripts
How to represent a script?
Could use proposition network for all the parts
… but maybe whole script should be a unit
Introspection suggests that it is activated as a unit
without interference from associated propositions
Experimental evidence (Bower, Black, Turner 1979)…
Got subjects to read a short story
Story followed a script, but didn’t fill in all details
They were then presented various sentences
Some from story, and some not
Some trick sentences were included:
Not from the story, but part of the script
Subjects were asked to rate 1(sure I didn’t read it) -7(sure I did read it)
Subjects had a tendency to think they read the trick sentences
Suggests that they activate the script and fill in the blanks in memory
…Starting to get a Model of the Mind
Propositional-schema representations stored in long-term
memory
Associative activation used to retrieve relevant memories
…but many details unspecified
Need more machinery to account for
Assess retrieved information, see does it relate to current goals
Decompose goals into subgoals
Draw conclusions, make decisions, solve problems
More importantly:
How to get new propositions and schemas into memory
Schemas are often generalised from examples, not taught
What about working memory?
Working Memory
Most long-term memory not “active” most of the time
Just keep a few things in working memory for current
processing
Very limited: try multiplying 3-digit numbers without paper
Working memory holds 3-4 chunks at a time
Why so limited? (it seems useful to have more nowadays)
Maybe complex circuitry required
Maybe costly in energy
Maybe tasks were less complex in environment of early humans
Or maybe more working memory would cause too many clashes,
or be too hard to manage
However limits can be overcome by skill formation
Note also: limit of 3-4 does not mean other “propositions” inactive
Could be a lot more going on subconsciously
Skill Acquisition
With a lot of practice we can “automate” many tasks
We distinguish this from “controlled processing” – using working memory
Once automated:
Takes little attention or working memory
(these are “freed up”)
Hard not to perform the task – cannot control it well
Most advanced skills use a combination
Automatic processes under direction of controlled processes, to meet goals
Examples: martial arts expert, or musician
Is Skill Acquisition Separate?
Evidence from Neuropsychology:
People with severe “anterograde amnesia”
Cannot learn new facts
i.e. can’t get them into long-term propositional memory
…but can learn new skills
Example:
Can learn to solve towers of Hanoi with practice
But cannot remember any occasion when they practised it
Suggests that a different part of the brain handles each
Skill may reside in visual and motor systems, rather than central systems
Maybe because of evolution:
Animals often have good skill acquisition
Maybe humans evolved a specific new module for high level functions
Mental Images
Sometimes we seem to evoke visual images in “mind’s eye”
Subjective experience suggests visual image is separate from propositions
…but need experimental evidence
In imagining a scene:
Example: search a box of blocks for 3cm cube with two adjacent blue sides
Properties are added to a description
But not so many properties as would be present in a real visual scene
Support, illumination, shading, shadows on near surfaces
Image does not include properties not available to visual perception
Other side of cube
Intuition suggests that “mind’s eye” mimics visual perception
Maybe it uses the same hardware?
Would mean that “central system” sends information to vision system
Mental Images
Hypothesis: there is a human “visual buffer”
Short-term memory structure
Used in both visual perception and “mind’s eye”
Special features/procedures:
Can load it, refresh it, perform transformations
Has a centre with high resolution
Focus of attention can be moved around
Assuming it exists… what good is it?
Allows you to pull things out of your visual long term memory
Use it to build a scene, with all spatial details filled in
Useful to plan a route, or a rearrangement of objects
Experiment: how many edges on a cube?
(Assuming answer is not in long term memory)
Experiments to show Mental Images
Test a special procedure: mental rotation
Experiments to show Mental Images
Time taken depended on how much rotation was needed
Suggests that we really rotate in the “visual buffer”
Experiments to show Mental Images
Experiments to show Mental Images
However… just because we rotate stuff doesn’t necessarily mean that we do
it in the “visual buffer”
…Need more evidence
PET brain scans have shown that the “occipital cortex” is used
“occipital cortex” is known to be involved in visual processing
So far…
The “Symbolic” Approach to
explaining cognition
an alternative…
the “Connectionist” approach…
So far…
The “Symbolic” Approach to
explaining cognition
an alternative…
the “Connectionist” approach…
Connectionist Approach
What is connectionism?
Concepts are not stored as clean “propositions”
They are spread throughout a large network
“Apple” activates thousands of microfeatures
Activation of apple depends on context, no single dedicated unit
Neural plausibility
Graceful degradation, unlike logical representations
Cognitive plausibility
Could explain entire system, rather than some task in central system
(symbolic accounts can be quite fragmented)
Could explain the “pattern matching” that seems to happen everywhere
(for example in retrieval of memories)
Explain how human concepts/categories do not have clear cut definitions
Certain attributes increase likelihood (ANN handles this well)
But not hard and fast rules
Explains how concepts are learned
Adjust weights with experience
Another Perspective on Cognitive Science / AI
We have seen multiple models for the mind,
and each has an “AI version” too
Propositions AI’s logic statements
Scripts AI’s case based reasoning
Mental images AI: some work, but not much
Connectionist models AI’s neural networks
This gives us another perspective on Cognitive Science / AI
Both are working in different directions
AI person starts with a computer and says
How can I make this do something that a mind does?
May take some inspiration from what/how a mind does it
Cognitive Science person starts with a mind and says
How can I explain something this does, using the “computer metaphor”?
May take some inspiration from how computers can do it
Especially from how AI people have shown certain things can be done
Another Perspective on Cognitive Science / AI
We have seen multiple models for the mind,
Which
modeltoo
is correct?
and each has an “AI
version”
Propositions AI’s logic statements
…possibly… all of them
Scripts AI’s case based reasoning
Mental images AI: some work, but not much
i.e. all working together
Connectionist models AI’s neural networks
This gives us another
perspective
on Cognitive
e.g. we have
seen that logic
could be Science / AI
Both are workingimplemented
in different directions
on top of Neurons
(need
in “clean” and
symbolic
AI person starts
withnota be
computer
saysway)
How can I make this do something that a mind does?
This
would give
opportunity
logical
May take some
inspiration
from
what/howfor
a mind
does it
Cognitive Science personreasoning,
starts with a mind and says
while still having “scruffy” intuitions
How can I explain going
something
does, using the “computer metaphor”?
on in this
the background.
May take some inspiration from how computers can do it
Especially from how AI people have shown certain things can be done