Demo Video

Video - high level details

• aim for about 4 minutes
• just do sensitivity case, talk about other cases in the summary if need to broaden the story

uil demo screen shots
screen shots as per 01_08_2009
annotated screen shots as per 01_08_2009

scenes and timings

Storyboard

1. Context
   • 25 learning components
   • predict meeting invitation from email, which project a web page is associated with
   • each one needs to learn in the wild, standalone without engineer after deployment
   • before deployment, design to ensure work well
   • goal - allow user to define new learning tasks
   • taking learning in the wild to a new level where the definition of the learning task itself is done in the wild
2. In this video
   • in this video we're going to show you an ex of how a user could teach calo a new learning task related to email management
   • suppose the user wants calo to learn to predict when it is appropriate to mark an email as sensitive.
   • the user conveys this to CALO by first writing an email that they would consider sensitive, and set the sensitivity field.
   • calo captures these steps as a procedure and then allows the user to introduce a branch at the point where they want the system to learn: under some conditions the user would want to mark the email as sensitive, under other conditions not. CALO will learn the classifier to distinguish these cases.
3. The demo
   1. Demonstrate creating email and adjusting sensitivity
   2. Edit the learned procedure
      • (Amend Learning Point button text?)
   3. Guidance UI
      • calo has to have some understanding of actions and properties that relate to it
      • ...
      • it infers that the most relevant thing at this point is the email being composed and the things relevant to it
      • from the calo kb it knows that the email has a sender, recipient, etc and all these things are potentially relevant
      • calo analyzes the procedure and figures out what attributes are in play at time of prediction and infers the label.
      • Here's a ui which describes email and it's properties
        we give the user an opportunity for the user to highlight things for calo to try first
      • standard statisitical learning requires lots of training to learn an accurate predictor. User has opportunity to focus CALO's attention on the aspects that the user thinks would be important in helping predict. This reduces the training required.
   4. Capturing a training example
      • CALO is able to recognize which branch is being taken in this email creation procedure, and so is able to collect training examples
   5. //time passes//
   6. Showing a case where the user forgets
      • Once CALO reaches a certain predictive accuracy, it will tell the user that it thinks that they forgot to set the sensitivity
      • Show dialog popping up in outlook if user forgets.

Issues with practice video

• email data sat Inbox has title "Two Weeks Ago". We'll want to drive some emails into the Charles Canyon account so the account doesn't appear inactive.
• clock time should not decrease as video moves forward
• we should choose a static background app (or no app) to be in play when appropriate (so that it does not appear that the background inexplicably changed while a UI was up in front)
• Question for Tom - Jed put a few seconds of coverage of the mouse showing events coming into the ITL UI, prior to composing an email. Should this be omitted?
• Question for Tom - Jed thinks we should employ fades in places - do you agree and if so, where?

1. Summary