Jesse English and Sandor Dornbush

1
Jesse English and Sandor Dornbush
With help from Dr. Zary Segall Chad Eby.
2
Activity Awareness

• Feedback loop between the user and their music.
• Use the users physiological state to influence
  the music selection.
• The music influences the users mood and mindset.

3
Proposed XPod Form Factor
4
Proposed XPod User Interface
5
Potential XPod Platform

• Nokia 5500 Sport Phone
• Embedded Accelerometers
• MP3 Player

http//europe.nokia.com/A4213113
6
Experimental XPod

• Bodymedia Device
• Tablet

7
Music and Human Activity

• Music has a strong connection with human
  activity.
• People like to play music that is appropriate for
  their current activity.
• Music will can affect a persons mental state and
  physical activity.

8
Music Selection

• Choose the song that the user would most likely
  like in this state.
• Use the user prior behavior to learn their
  preferences.
• Combine a star rating system with body
  information.

9
Song Selection
Choose a Song Randomly
Learning Algorithm Estimate Preference
Flip coin weighted by the Estimated preference
Tails
Heads
Play Song
10
Results
11
Results

• XPod creates a custom context aware playlist.
• The addition of state information can improve the
  accuracy of the learning system.
• Prefrence(SongTime,Activity) ? Preference(Song)
• Neural Network seems to be best.
• Generalizes best
• Over trains very quickly
• SMO gets the exact.
• Does not generalize well.

12
Network Structure

• 289 Input neurons
• 1 Output neuron
• 1-50 Hidden neurons
• 4 is the best

13
Future Work

• Develop a prototype device.
• Nokia 5500 Sport
• iPaq and Backpaq
• Incorporate song metadata
• Human Generated, eg Last.fm
• Machine Generated
• Incorporate other meta information
• Location Information
• Recent Phone Calls
• Weather

14
Questions?