Small Group Subproblems

From Noisebridge
(Difference between revisions)
Jump to: navigation, search
Line 24: Line 24:
 
*Noisebrain(Jared, Mike)
 
*Noisebrain(Jared, Mike)
 
**Install sensors all over Noisebridge.** (need to deal with some privacy concerns first)
 
**Install sensors all over Noisebridge.** (need to deal with some privacy concerns first)
***IR? PIR? Visual spectrum? Sonar? Decibel level? Audio seems like privacy landmine.
+
***IR? PIR? Visual spectrum? Sonar? Decibel level?
 +
***Any logging of raw Stills, Video or Audio seems like privacy landmine.  That said, I suspect aggregating it in a mathy way should be OK with folks.
 
**We can train it to learn about what's going on in Noisebridge, for example:
 
**We can train it to learn about what's going on in Noisebridge, for example:
 
***How many people are in the space?
 
***How many people are in the space?
Line 33: Line 34:
 
***Feed this information into [[Noise-Bot]] in addition to it's on-board information.
 
***Feed this information into [[Noise-Bot]] in addition to it's on-board information.
 
***Could provide additional context for Music Recommendation project below.
 
***Could provide additional context for Music Recommendation project below.
***Neat widgets for Noisebridge website showing what's up.
+
***Neat dashboard widgets for Noisebridge website/wiki showing what's going on.
 +
***Hack the lighting in the various parts of the space to be auto controlled.
 
*Music Recommendation System (Jared)
 
*Music Recommendation System (Jared)
**Given different people in a space (who voluntarily "checkin"), and a jukebox, play music that appeals to their tastes.
+
**Given different people in a space (who voluntarily "checkin"), and a jukebox, play music that appeals to their tastes.  Could also be used for personal playlist creation.
**<strike>Hard part is data gathering</strike> We are just use last.fm [http://www.last.fm/download#content_ipod scrobbler] & [http://ws.audioscrobbler.com/1.0/user/jareddunne/recenttracks.rss rss feeds].
+
**I would like to take a different approach to the recommendation algorithm from my undergrad approach, which was based around domain specific knowledge about relationships and influences about various musicians.  Instead I'd like to train models using features like time of date, location, weather, etc tied to track plays by individual users.  I do still want past play history to also influence the weightings in some way.
 +
**Capturing music play events isn't hard: We are just use last.fm [http://www.last.fm/download#content_ipod scrobbler] & [http://ws.audioscrobbler.com/1.0/user/jareddunne/recenttracks.rss rss feeds].
 +
**Weather and Location feature collection might be some what time consuming but not unmanageable.
 +
**When a music play event (a 'scrobble') happens, we'd collect and log all the contextual information possible.
 +
***When a song is scrobbled on iPhone or iTunes, we'd know about the event almost instantly allowing more contextual information.
 +
***Songs played on normal iPods get scrobbled when it's synced with iTunes with scrobbler plugin.  This delay might make collecting some contextual features difficult after the fact.  Nonetheless the partial information is still useful.  
  
 
TODOs:
 
TODOs:
 
*Put together a github project
 
*Put together a github project
 +
**Maybe we should just put things in the existing noisebridge-ml github repo as subfolders?
 
*Key is to have a bunch of projects going on at the same time so people can work on whatever they like at a given time.
 
*Key is to have a bunch of projects going on at the same time so people can work on whatever they like at a given time.

Revision as of 01:24, 17 February 2011

Joe H proposed that we come up with some smaller problems that allow people to interactively solve them in short periods of time in ML meetups. We can construct a git hub project for them, create a skeleton project, and use issue tracking to list the problems and work on them. Here we list these problems and plan to build them out.

  • Time Series Data (Erin)
  • Rock Climbing (Joe H)
    • Rock climbing data analysis
    • Generate predictive system for how much someone should climb based on how they're doing that week.
    • Noted that amount of deep sleep can be correlated to mood, which can be correlated to climbing.
    • Suggestion: we all track one aspect of our daily routines and combine it into one big dataset.
  • 4square Checkins (Joe H)
    • Use 4square checkins to find out what would be the best time to visit a place in a region.
  • Education Learning Problem (Tom)
    • Determine how people learn over time.
    • Figure out how difficult questions are, what kind of knowledge they test
    • Determine how good are people at certain types of knowledge.
  • Auditory Brain Cells (Mike S)
    • Soon to be posted data on crcns.org of auditory brain cells.
    • Figure out how to predict brain cell output.
  • Deep Belief Networks for Images/Sounds (Mike S)
    • Train deep belief networks on images, project output to monitor to see what network is "thinking"
    • Do same for sounds
    • Use auto-encoders to recall memories based on new input
  • Kinects (Jared, Praveen)
    • Get another kinect and do awesome shit with it.
  • Noisebrain(Jared, Mike)
    • Install sensors all over Noisebridge.** (need to deal with some privacy concerns first)
      • IR? PIR? Visual spectrum? Sonar? Decibel level?
      • Any logging of raw Stills, Video or Audio seems like privacy landmine. That said, I suspect aggregating it in a mathy way should be OK with folks.
    • We can train it to learn about what's going on in Noisebridge, for example:
      • How many people are in the space?
      • What sort of activities are happening? Classes? Working? Hacking? Construction?
      • What parts of the space are being used? Main Hall? Church? Turing? Kitchen?
      • Where is there motion? Where are things out of place?
    • Application ideas for the Noisebrain once working:
      • Feed this information into Noise-Bot in addition to it's on-board information.
      • Could provide additional context for Music Recommendation project below.
      • Neat dashboard widgets for Noisebridge website/wiki showing what's going on.
      • Hack the lighting in the various parts of the space to be auto controlled.
  • Music Recommendation System (Jared)
    • Given different people in a space (who voluntarily "checkin"), and a jukebox, play music that appeals to their tastes. Could also be used for personal playlist creation.
    • I would like to take a different approach to the recommendation algorithm from my undergrad approach, which was based around domain specific knowledge about relationships and influences about various musicians. Instead I'd like to train models using features like time of date, location, weather, etc tied to track plays by individual users. I do still want past play history to also influence the weightings in some way.
    • Capturing music play events isn't hard: We are just use last.fm scrobbler & rss feeds.
    • Weather and Location feature collection might be some what time consuming but not unmanageable.
    • When a music play event (a 'scrobble') happens, we'd collect and log all the contextual information possible.
      • When a song is scrobbled on iPhone or iTunes, we'd know about the event almost instantly allowing more contextual information.
      • Songs played on normal iPods get scrobbled when it's synced with iTunes with scrobbler plugin. This delay might make collecting some contextual features difficult after the fact. Nonetheless the partial information is still useful.

TODOs:

  • Put together a github project
    • Maybe we should just put things in the existing noisebridge-ml github repo as subfolders?
  • Key is to have a bunch of projects going on at the same time so people can work on whatever they like at a given time.
Personal tools