48
Ensemble Methods

Ensemble Methods. “No free lunch theorem” Wolpert and Macready 1995

Embed Size (px)

Citation preview

Page 1: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Ensemble Methods

Page 2: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

“No free lunch theorem” Wolpert and Macready 1995

Page 3: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

“No free lunch theorem” Wolpert and Macready 1995

Solution search also involves searching for learners

Page 4: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Different algorithms

Page 5: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Different algorithmsDifferent parameters

Page 6: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Different algorithmsDifferent parametersDifferent input

representations/features

Page 7: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Different algorithmsDifferent parametersDifferent input

representations/featuresDifferent data

Page 8: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Base learner

Page 9: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Diversity over accuracy

Page 10: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Model combination

Page 11: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

VotingBaggingBoostingCascading

Page 12: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995
Page 13: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995
Page 14: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995
Page 15: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Data set = [1,2,3,4,5,6,7,8,9,10]

Samples: Input to learner 1 = [10,2,5,10,3] Input to learner 2 = [4,5,2,7,6,3] Input to learner 3 = [8,8,4,9,1]

Page 16: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Create complementary learners

Page 17: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Create complementary learnersTrain successive learners on the

mistakes of predecessors

Page 18: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Weak learners combine to a strong learner

Page 19: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995
Page 20: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995
Page 21: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995
Page 22: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995
Page 23: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Adaboost – Adaptive Boosting

Page 24: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Adaboost – Adaptive BoostingAllows for a smaller training set

Page 25: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Adaboost – Adaptive BoostingAllows for a smaller training setSimple classifiers

Page 26: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Adaboost – Adaptive BoostingAllows for a smaller training setSimple classifiersBinary

Page 27: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Modify probability of drawing examples from a training set based on errors

Page 28: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995
Page 29: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995
Page 30: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995
Page 31: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995
Page 32: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

α1= 12log(

1− error

error)

α1= 12log(

1− .33

.33)

α1= 0.35€

error = 0.33

Step 3

Page 33: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995
Page 34: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995
Page 35: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995
Page 36: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Demo

Page 37: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Sequence classifiers by complexity

Page 38: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Sequence classifiers by complexityUse classifier j+1 if classifier j

doesn’t meet a confidence threshold

Page 39: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Sequence classifiers by complexityUse classifier j+1 if classifier j

doesn’t meet a confidence thresholdTrain cascading classifiers on

instances the previous classifier is not confident about

Page 40: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Sequence classifiers by complexityUse classifier j+1 if classifier j

doesn’t meet a confidence thresholdTrain cascading classifiers on

instances the previous classifier is not confident about

Most examples classified quickly, harder ones passed to more expensive classifiers

Page 41: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Boosting and Cascading

Page 42: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995
Page 43: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995
Page 44: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995
Page 45: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995
Page 46: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995
Page 47: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Object detection/trackingCollaborative filteringNeural networksOptical character recognition ++BiometricsData mining

Page 48: Ensemble Methods.  “No free lunch theorem” Wolpert and Macready 1995

Ensemble methods are proven effective, but why?