The Future of Mobile
Applications
John Canny
UCB EECS
Marc Davis
UCB School of Information
& Yahoo! Research Berkeley
The Business
• There are 6.5 billion people on earth
- only about 1.2 billion in “developed” countries
• They will buy 800 million mobile phones this year
- one person in eight on the planet
800m
• That’s 4x PC or TV unit sales
• Fraction of smartphones
should reach 40% by 2009
- most common “computer”
200m
What kind of computer is it?
This year’s Smartphone (free with service contract)
• 150-200 MHz ARM processor
• 32 MB ram
• 2 GB flash (not included)
Windows-98 PC that boots quickly!
Plus:
200 mips
• Camera
• AGPS (Qualcomm/Snaptrack)
• DSP cores, OpenGL GPU
• EV-DO (300 kb/s), Bluetooth
What’s Coming
In the past, the platform was driven by voice+messaging
Now the high end is driven by video, gaming, location,…
The result is diversification of the platform, and sudden
jumps in performance, e.g. Qualcomm has 4 platforms:
1. Value platform (voice only)
PC
2. …
Log P
3. …
4. Convergence platform (MP3 player,
gamer, camera,…) several times
the performance of today’s high-end
time
The Inevitable…
In response to MIT’s $100 laptop, Microsoft last month
proposed the cell phone computer for developing countries:
Bollywood
on demand
click here
Back to the future, which is…
Using Context
• Location, time, BT neighborhood,…
• Community
• User History
Harnessing Content
• Text, Images, Video + Metadata
• Speech Recognition
• Computer Vision
What’s wrong today…
Did you ever try to find a neighborhood restaurant using a
mobile browser…
and find it while you were in the same neighborhood?
In a car you might end up in the next county…
Luckily a house stopped this
driver before they got into serious
trouble.
Context-Awareness
Context-awareness is the holy grail for next generation
mobile applications:
• Location (e.g., video store) heavily shapes
the user’s likely actions.
• The system can present streamlined
choices – “here are your top-10 video
suggestions with clickable previews”.
• For users this is very convenient.
• For vendors,…
Context-Awareness and Pro-Activity
Knowledge of user background and context provide great
opportunities for pro-active services:
• “It’s 7pm and you’re in San Francisco, would you like me to
find a nearby restaurant?”
Context-Awareness and Pro-Activity
Knowledge of user background and context provide great
opportunities for pro-active services:
• “It’s 7pm and you’re in San Francisco, there is a table
available two blocks away at Aqua. Would you like me to
book it?”
Context-Awareness and Pro-Activity
Knowledge of user background and context provide great
opportunities for pro-active services:
• “It’s 7pm and you’re in San Francisco, there is a table
available two blocks away at Aqua, and they have a special
on Salmon in parchment for $28. Would you like me to
book a table, and order the special?”
Context-Awareness and Recognition
Consider now a speech recognizing version of this
application:
• “It’s 7pm and you’re in San Francisco, there is a table
available two blocks away at Aqua, and they have a special
on Salmon in parchment for $28. Would you like me to
book a table, and order the special?”
User: Yes or No
Context-Awareness and Activity
People’s actions are part of larger wholes called activities.
When you plan an evening out it may include:
• Going for coffee
• Seeing a movie
• Eating dinner
- planned and coordinated by the phone
• Sharing photos on your cameraphone
• Scoring your date in real-time *
It’s a social platform!
Activity-Based Design
Activity-based design creates chains of services (BigTribe) or
menus of related actions. More examples:
• Planning a trip: hotel, car, events
• Going back to school: housing, books etc.
• Shopping for holiday gifts
• Moving house
• Hobbies: Needlepoint,…
Monster truck racing
Some of these services exist, but activity analysis supports
automatic discovery and customization of them.
Sociotechnical Systems
• The problems of the internet are not purely
technological
• Need to conduct sociotechnical analysis and
design of large scale internet systems and
applications at the intersection of media,
technology, and people
• Leverage media metadata created by contextaware devices, content analysis, and
communities
Signal-to-Symbol Problems
• Semantic Gap
– Gap between lowlevel signal analysis
and high-level
semantic
descriptions
– “Vertical off-white
rectangular blob on
blue background”
does not equal
“Campanile at UC
Berkeley”
Signal-to-Symbol Problems
• Sensory Gap
– Gap between how an object appears and what it is
– Different images of same object can appear
dissimilar
– Images of different objects can appear similar
Computer Vision and Context
•
•
•
•
•
You go out drinking with your friends
You get drunk
Really drunk
You get hit over the head and pass out
You are flown to a city in a country you’ve never been to with
a language you don’t understand and an alphabet you can’t
read
• You wake up face down in a gutter with a terrible hangover
• You have no idea where you are or how you got there
• This is what it’s like to be most computer vision systems—
they have no context
• Context is what enables us to understand what we see
Campanile Inspiration
MMM: Mobile Media Metadata Idea
• Leverage the spatio-temporal context and social
community of media capture in mobile devices
– Gather all automatically available information at the point
of capture (time of capture, spatial location, collocated
phone users, etc.)
– Analyze contextual metadata and media to find similar
media that has been captured before
– Use patterns in previously captured media/metadata to
infer the content, context, and community of newly
captured media
– Interact with users to augment system-supplied metadata
for captured media
MMM: Mobile Media Metadata Projects
Mobile Media Metadata
Davis, Canny, et al.
UC Berkeley
Context-Aware Face Recognition
Context-Aware Face Recognition
• Face recognition alone
- 43% accurate
(state of the art computer vision)
• Context analysis alone
- 50% accurate
(Face prediction from
contextual data on the phone)
• Context+Content analysis
- 60% accurate
Context-Aware Place Recognition
• Image analysis alone
- 30% accurate
• Context analysis alone
- 55% accurate
• Context+Content analysis
- 67% accurate
MMM2: Context to Community
Photo Share Guesser
0.8
0.7
0.6
0.5
SFA prediction
Baseline
0.4
0.3
0.2
0.1 1
2
3
4
5
6
7
8
9
10
Photo Level of Interest (LOI) Browser
PhotoCat: Context-Aware Photo Browser
Technologies
We have been developing core technologies for context and
content mining for the last 5 years:
• Accurate, scalable personalization (used in MMM2)
• Algorithms for integration of personal and context
information, and for activity discovery
• Methods to preserve privacy while mining user location
history and online behavior
We’ve also worked with a company (BigTribe) through 3
funded NSF SBIRs, to migrate these ideas into products.
Harnessing Large, Mixed Content Sources
Early access to an XML Content-Base (Mark Logic CIS)
We built an efficient location+metadata server from diverse
well- and poorly-structured data sources.
• Street data comes from XML Census data (Tiger/GML).
• Restaurant data is from the Open
Directory, partly structured.
• Addresses converted to
LAT/LONG by the database.
• The map you see is produced
entirely using XQuery (SVG).
Picante Playa
Location Content-Base
A native XML engine supports
efficient tree traversal.
The location C-B uses R-tree
organization as its XML schema
The result is that our “software” spatial
database has the same efficiency as
custom spatial databases.
i.e. not only are data types extensible,
but also the types of query that are
efficiently supported.
Context-Aware Design – Glaze
Designing new context-aware apps works best “in the wild.”
We are doing a participatory design experiment with 20
AGPS phones this spring.
Users carry the phones with them everywhere, be able to use
some seed applications, and otherwise create their own
“micro-apps” through noun-verb composition.
Perceptual Interfaces - Vision
We needed continuous mouse input for map browsing, so we
developed TinyMotion, a software mouse for cameraphones.
By moving the camera against any background, real-time
image motion estimation provides mouse coordinates.
Also great for games – demo in BID lab
Perceptual Interfaces - Vision
Cameraphones are capable of much more. Right now, the
vision algorithms available include:
• Motion
• Barcodes
• OCR text (business cards etc.)
Coming soon:
• Face recognition
• Building or streetscape recognition
Perceptual Interfaces - Speech
Speech recognition technology has improved steadily in the
last ten years, particularly in noisy environments.
Speech was never a good match for office environments.
But the mobile playing field is completely different.
Mobile users often need their eyes and hands free, and the
phone will always have a voice channel for telephony.
Speech on Mobile Phones
Restricted speech recognition is available on many phones.
Large-vocabulary recognition just appeared on cell phones
last year (Samsung P207). Its a huge step. It enables the
next generation of mobile speech-based apps:
• Message dictation
• Web search
• Address/business lookup
• Natural command forms
(no need to learn them)…
Most of this technology was developed
in the US by VoiceSignal Technologies.
Research in Mobile Speech
We are developing a state-of-the-art (continuous acoustic
model) recognizer for SmartPhones.
The goals are:
• To provide an open platform for next-generation, speechbased interfaces on mobile devices.
• To support integration of contextual knowledge in the
recognizer.
• To allow efficient exploration of the higher levels of dialogbased interfaces.
Speech for Developing Regions
Speech is an even more important tool in developing regions.
Literacy is low, and iconic (GUI) interfaces can be hard to use.
Unfortunately, IT cannot help most of these people because
they lack even more basic skills – fluency in a widely-spoken
language like English or Mandarin.
This project focuses on teaching English
in an ecologically appropriate way.
Speech-based phones are ideal for this.
Speech for Developing Regions
Speech (with headset) allows students to learn while working.
It leaves their eyes and hands free, and engages their minds
during tedious, manual work.
Some game motifs:
• Safari: hear sound & say the name in English
• Karoake: in English
• Listen and summarize: BBC, cricket etc.
• Treasure hunt: leave LB clues in English
• Adventure games: dialog-driven scenarios
In Summary,
The Future of Mobile is:
• Using Context
• Harnessing Content
• Context  Proactivity
• C/P  Content sharing
• Perceptual: Speech/Vis
• 5 Billion new users
Upcoming…
Special issue of ACM Queue magazine on context-aware and
perceptual interfaces (summer 06?) JFC guest Ed.
Workshop on Mobile Applications
Planning an event on campus later this semester.
Send mail to jfc@cs.berkeley.edu if interested.
Demos and Posters Today
You can see the projects discussed here in the BID lab
(Berkeley Institute of Design) open house, 2-4pm:
• Tinymotion camera mouse
• Glaze location service design
• Speech recognition for cell phones
• English-Language learning with cell phones
+ a dozen other projects
The lab is in 354-360 Hearst Mining Bldg.
Acknowledgements
Thanks to the sponsors of this work:
and..