Project Oxygen: A Retrospective

1
Project Oxygen A Retrospective

• Larry Rudolph

2
What was it
3
Lessons

• Human-centric input inherently ambiguous
• goal is to disambiguate
• chop into pieces with meaning
• Stepford Wives
• input --gt recog pieces --gt reassembly --gt output

4
Lessons

• information must flow across abstraction barriers
• app --gt what is expected
• n-best approximation --gt app

5
Lessons

• State of the world is stochastic
• most of the time, stuff isnt working correctly
• Multi-modes compensate for errors

6
Lessons

• Lots of similarities in diverse fields
• Bayesian inference widely used
• speech, vision, gestures, handwriting, location,
  congestion control
• Lots of app-specific optimizations
• real engineering work, not easily generalized
• Its all the same, but different

7
Lessons

• Location (indoor) is key
• Hard to beat commodity items
• two heads better than one

8
Things done right (usually by accident)

• No one overall architecture
• No major standardization effort
• Lots of pairwise interactions
• Fewer constraints with state-of-the-art systems
• Professional videos
• very little cheating real deadlines
• Publications in subfield areas -- measurable
  progress

9
Version 2.0?

• Eat our own sausage
• lots of support staff for internal technology
  transfer
• speech recognition on every desk
• free-hand sketching --gt powerpoint
• comprehension-lite versions
• ala speech builder

10
Brittleness

• up and running all the time (not just during
  demos)
• too many publish subscribe mechanisms
• must be able to measure

11
Scenarios

• too much emphasis
• for every wouldnt it be cool scenario, there
  exists a nightmare scenario
• there is no cool-meter or cool-metric

12
Metric for Success of Oxygen?

• did it change the world? No
• did it help the world change? Yes