Transcript Dr Russell Beale`s Guest Lecture on AI IN HCI

advancedinteractiongroup
Synergistic Serendipity
- or how to make good things happen
Russell Beale
1
advancedinteractiongroup
Overview
• Systems should facilitate human activities
–
–
–
–
communication
learning, problem solving
work, play
love, life
• Use intelligence in interactive systems
– AI to aid HCI
– Intelligent design
2
advancedinteractiongroup
Talk summary
• Synergy of intelligence and design
• Three example systems
– Data mining - keeping it simple
– Supportive browsing - looking into your future
– Social interaction - dating for the socially inept
3
advancedinteractiongroup
Interaction
• Interaction with technologies should not be easy
• Why? Because the technologies should be incidental
to the task, assisting it not obstructing it
– Example
• BMW 5 series (old model) - more computing power than used
to put man on the moon
• Interaction?
– Get in, turn key, drive
• Engine management system highly complex, but we don’t
interact with it directly
4
advancedinteractiongroup
Approach
• Introduce synergistic intelligence into appropriate
devices and interactions
• Design interaction with awareness of aesthetics and
user
• Why so useful?
– Most computing systems have vast numbers of unused
compute cycles – put them to good use
– More happening behind the scenes, need appropriate
information
– Users can contribute hugely to interaction success, if you let
them
– aiming for magic
5
advancedinteractiongroup
Magic interaction
• What is magic?
– Amazing
– achieving what you thought was impossible
– if you think about it, a perfectly ordinary
explanation, but you don’t really want to know, it’s
just great it does it...
6
advancedinteractiongroup
How?
• Understand people, technology, and
interaction
– e.g. temporal modelling of communication
systems with owned triggers
• Map technical achievements onto high-level
needs
• Design
– user-centred for needs-meeting
– bold (but tested) for new systems
7
advancedinteractiongroup
Interaction framework
input input
intelligence
user
system
output output
8
advancedinteractiongroup
Synergistic intelligence
• Intelligent media systems should be synergistic
• The whole should be greater than the sum of the
parts
• We should benefit from their capabilities
• They should benefit from ours
• How?
– Develop their knowledge over time
– synergy between human abilities (vision, problem solving,
new situations) and machine skills
9
advancedinteractiongroup
Not Artificial Intelligence
• Goal of artificial intelligence - to make
computers behave like they do in the
movies
• Silicon humans - but why?
• Better to have Appropriate Intelligence:
the right level of machine ability
• Best to have Synergistic Intelligence:
combination of machine and user
becomes much more powerful
10
advancedinteractiongroup
Three projects
•
•
•
Synergistic data mining
Supportive browsing
Social interaction
11
advancedinteractiongroup
1. Synergistic data mining
• Data mining: the search for something
interesting
• But what is interesting?
• Interesting varies – is it the fact that things
seem to follow a straight line – or is it the odd
outliers from that line?
25
20
15
Series1
10
5
0
0
5
10
15
20
25
12
advancedinteractiongroup
Designing for results
• Good design maximises the benefits of the different
systems
• Users
– Can define interesting
• If they can see the data
– Are good at spotting clusters, trends
• Human visual system excellent in presence of occlusion, noise,
etc.
• Computers
– Can process large quantities of data
– Can follow complex, time-consuming algorithms
13
advancedinteractiongroup
Synergistic data mining
• Design system to visualise
abstract data
• Place data points randomly in
space
• Connect them with springs
QuickTime™ and a
TIFF (LZW) decompressor
are needed to see this picture.
– Forces dependent on similarity
between items
• Evolve system to stable state
– “Similar” things will be close
together
– Dissimilar things far apart
– Unconnected things separated
14
advancedinteractiongroup
Generating rules
• Now we have visualisation, need to explain it
• Ask computer to use a.i. approaches to
generate some explanations
– Statistics?
– Neural networks?
– Rules?
• But not long complex ones
15
advancedinteractiongroup
People and Information
• Users can understand simple rules
• Machines tend to focus on accuracy at the
expense of simplicity
• Need to generate simple rules at first, then
become more complex as we go on
16
advancedinteractiongroup
Evolving simple rules
• Rules are of the sort
IF colour = red & texture= soft & size < 3.2 THEN fruit = strawberry
• Easiest to see the wood, then the trees
– General, less accurate rules
– Then refine to detail
• Use symbolic GA to refine rules
– Fitness defined by accuracy, coverage, etc.
– And size of rule
– Can tune system to produce level of detail required
17
advancedinteractiongroup
Evolving good rules
• Use a genetic algorithm approach
• How it works
1. Create a random set of rules that may or may not describe
the data
2. See how well they do
3. Breed the best ones
Take bits from each (crossover)
Change bits of them (mutation)
4. Repeat from 2 until you get a good answer
18
advancedinteractiongroup
Evolving useful rules
• Check out step 2.
– “see how well they do”
• If we make short, easy to understand rules
score well, then the system will favour these
• Gives us an overview
• Then can reduce the impact of length of the
scoring, increase the demands for accuracy
• Gives us detail
19
advancedinteractiongroup
Example
• Detail
– If Ferrari red, then fast OR if red and 3 litre engine,
then fast, OR if 2 litre engine and not Fiat built
before 1998 then fast, OR if Subaru and import
and rally special and not estate then fast – 99.7%
accurate
• Short overview
– “all fast cars are red” – 85% accurate
20
advancedinteractiongroup
Visualising the rules
• Add the rules back into the visualisation as
large nodes, linked to the data they cover
• Can identify coverage, accuracy,
misclassifications, multiple classifications,
redundant rules etc.
21
advancedinteractiongroup
Feedback
• Users define what is “interesting” or
merits further investigation
• System describes it
– Classification, description
• Visualises results
• Iterate
QuickTime™ and a
TIFF (LZW) decompressor
are needed to see this picture.
22
advancedinteractiongroup
2. Supportive browsing
• Analyse internet usage
– three types: searching, monitoring, browsing
– supported by Google, RSS, web browser
• Browser not user centred, not task related
• Typical problems
– Browsing interesting topic but taken off it
– Seen something recently but can’t relocate it
23
advancedinteractiongroup
www
Mitsikeru
• Proxy to capture and model user
interests over time, cluster into concepts
• Proxy look-ahead for current page
• Analyse pre-fetched pages for relevant
material
• Annotate current page with ‘post-it’ style
notes when hover over links
24
advancedinteractiongroup
Interesting…
• Interesting pages
– Contain words that are
more common in personal
and current browsing than
in general use
– Especially if highly unusual
words
– Use Bayesian statistics to
calculate values
25
advancedinteractiongroup
Unobtrusive intelligence
•
•
•
•
Display results as ‘expanded tooltips’
Requires no additional work by user
Provides peek into the future
Allows you to browse as usual but with better
idea of where you’re going
• Cross-browser compatibility
26
advancedinteractiongroup
3. Social and contextual
interaction
• People are inherently social
• Can be a problem for people to meet other likeminded people to develop relationships
– “my friend says, do you want to go out with her?”
– can we support this?
• Use bluetooth and phone to advertise own interests
and desires
• When relevant interesting person is nearby, systems
check each other out and then notify their owners...
27
advancedinteractiongroup
Further social interactions
• Filesharing
• Jokes
• Gossip
– All oil the wheels of group dynamics
• Mobile blogging
• Context-aware systems
– Understand context though model
– Context like a movie
– CAGE
28
advancedinteractiongroup
Summary
• Understanding and theory + intelligent design +
appropriate intelligence = great systems
• 3 examples
– All show aspects of above in different amounts of detail
– All different
• Application domain, effects, technologies
– All similar
• Same principles applied
• Support the user in natural interactions
• Make the technology magic
29