Machine Learning for the Sensor Revolution

Transcript

1. Berlin     November  2013   Machine  Learning  for  the   Sensor  Revolu=on   Henrik  Brink        @brinkar  
2. The  old  way  
3. The  recent  way  
4. The  new  way...  
5. ...requires  huge   data-­‐centers!  
6. ~  30  terabytes  per  day  
7. Find  the  needle  in  an   astronomical  haystack  
8. Supernova  discovered  11  hr  aLer  explosion  by  ML   algorithm.  Nearest  SN  Ia  in  >  3  decades!  
9. Brink  et  al.  2013   Machine  Learning  for   supernova  discovery   Real  vs  Bogus  
10. The  Internet  of  Things   “machine-­‐generated  data  will  con=nue  to   grow  as  fast  as  Moore’s  Law  lets  it”   “It  is  unrealis=c  to  hope  ever  to  keep  most  or   all  machine-­‐generated  data”   “most  data  (by  volume)  will  be  machine-­‐ generated”   hUp://www.dbms2.com/2010/01/17/three-­‐broad-­‐categories-­‐of-­‐data/  
11. Example:  Smart   meters         Predic=ng   energy  usage   Self-­‐learning   thermostat  
12. Example:  Smartphone  sensors   561  columns:  a   few  seconds  of   accelerometer  and   gyroscope  data   6  labels:  walking,   walking  upstairs,   walking  downstairs,   si^ng,  standing,  laying     98%  accuracy  with   nonlinear  ML.   Standing  and   si^ng  is  most   confused  
13. •  How  do  we  store  and  manage  the  data?     •  How  do  we  find  needle-­‐in-­‐the-­‐haystack   events?   •  How  do  we  make  decisions  in  real-­‐=me?   Challenges  
14. Big  Data     tech  is  now   commodity  
15. •  How  do  we  store  and  manage  the  data?     •  How  do  we  find  needle-­‐in-­‐the-­‐haystack   events?   •  How  do  we  make  decisions  in  real-­‐=me?   Challenges  
16. Supervised  Machine  Learning   Training   data   Labels   Features   Supervised   ML   algorithm   Predic=ve   model   New   data   Features   Predicted     label  for   new  data   Model   evalua=on  &   op=miza=on  
17. Linear  models   Disadvantage   Linear     Advantage   Scales  well  
18. Nonlinear   models   Smartphone  ac=vity   1%  false-­‐posi=ves   99%  detec=on  rate   Source:  Scikit-­‐Learn  classifier  comparison  
19. Random  Forests®   ü   Classifica=on,  regression      and  clustering   ü   Nonlinear,  parallelizable,    easily  tuneable   ü   Categorical,  floats,  integers,  boolean      (no  normaliza=on  needed)     ü   Missing  data,  imbalanced  training  data   Random  Forests®  is  a  registered  trademark  of  Salford  Systems  
20. ML  workflow   Data  inges=on   Model   deployment   Model  valida=on   Feature   engineering  
21. •  How  do  we  store  and  manage  the  data?     •  How  do  we  find  needle-­‐in-­‐the-­‐haystack   events?   •  How  do  we  make  decisions  in  real-­‐=me?   Challenges  
22. Real-­‐=me  machine  learning   Deep  learning  can  learn  to   recognize  humans  like   humans.       Also  cats.  
23. WiseRF™  –  Random  Forests®   Random  Forests®  is  a  registered  trademark  of  Salford  Systems   Predic=ng   8M  hand-­‐   wriUen   digits   Embedded  devices   and  sensors   Machine  Learning   applica=on  framework   Automated  data  science   Predic=ons  via  API  or   embedded  models   @wiseio              contact@wise.io  
24. Join  the  community   Data  Science  and  Machine   Learning  in  Copenhagen     100+  members  <  1  week   2500+  members     25  meetups  in  <  2  years  
25. Ques=ons?