Download presentation
Presentation is loading. Please wait.
Published byAllen Floyd Modified over 9 years ago
1
CSE 473 Ensemble Learning
2
© CSE AI Faculty 2 Ensemble Learning Sometimes each learning technique yields a different hypothesis (or function) But no perfect hypothesis… Could we combine several imperfect hypotheses to get a better hypothesis?
3
© CSE AI Faculty 3 Example this line is one simple classifier saying that everything to the left is + and everything to the right is - Combining 3 linear classifiers More complex classifier
4
© CSE AI Faculty 4 Analogies: Elections combine voters’ choices to pick a good candidate (hopefully) Committees combine experts’ opinions to make better decisions Students working together on Othello project Intuitions: Individuals make mistakes but the “majority” may be less likely to Individuals often have partial knowledge; a committee can pool expertise to make better decisions Ensemble Learning: Motivation
5
© CSE AI Faculty 5 Technique 1: Bagging Combine hypotheses via majority voting
6
© CSE AI Faculty 6 Bagging: Analysis Error probability went down from 0.1 to 0.01!
7
© CSE AI Faculty 7 Weighted Majority Voting In practice, hypotheses rarely independent Some hypotheses have less errors than others all votes are not equal! Idea: Let’s take a weighted majority
8
© CSE AI Faculty 8 Technique 2: Boosting Most popular ensemble learning technique Computes a weighted majority of hypotheses Can “boost” performance of a “weak learner” Operates on a weighted training set Each training example (instance) has a “weight” Learning algorithm takes weight of input into account Idea: when an input is misclassified by a hypothesis, increase its weight so that the next hypothesis is more likely to classify it correctly
9
© CSE AI Faculty 9 Boosting Example with Decision Trees (DTs) training case correctly classified training case has large weight in this round this DT has a strong vote. Output of h final is weighted majority of outputs of h 1,…,h 4
10
© CSE AI Faculty 10 AdaBoost Algorithm
11
© CSE AI Faculty 11 AdaBoost Example Original training set D 1 : Equal weights to all training inputs Goal: In round t, learn classifier h t that minimizes error with respect to weighted training set h t maps input to True (+1) or False (-1) Taken from “A Tutorial on Boosting” by Yoav Freund and Rob Schapire
12
© CSE AI Faculty 12 AdaBoost Example ROUND 1 MisclassifiedIncrease weights z 1 = 0.42
13
© CSE AI Faculty 13 AdaBoost Example ROUND 2 z 2 = 0.65
14
© CSE AI Faculty 14 AdaBoost Example ROUND 3 z 3 = 0.92
15
© CSE AI Faculty 15 AdaBoost Example h final sign(x) = +1 if x > 0 and -1 otherwise
16
Example 1: Semantic Mapping
17
Motivation Human-Robot interaction: User: “Go to the corridor” Room Corridor Doorway
18
Shape Room
19
Observations Room Doorway
20
Observations Room Corridor Doorway
21
Simple Features Gap = d > θ f = # Gaps Minimum f =Area f =Perimeter f = d d d i d f = d d Σ d i
22
Experiments Training (top) # examples: 16045 Test (bottom) # examples: 18726 classificatio n: 93.94% Building 079 Uni. Freiburg Room Corridor Doorway
23
Application to a New Environment Training map Intel Research Lab in Seattle
24
Application to a New Environment Training map Intel Research Lab in Seattle Room Corridor Doorway
25
Example 2: Wearable Multi- Sensor Unit © CSE AI Faculty Battery Camera (on ribbon cable) GPS receiver iMote2 + two sensor boards MicrophoneCamera Light sensors 2 GB SD card Indicator LEDs Records 4 hours of audio, images (1/sec), GPS, and sensor data (accelerometer, barometric pressure, light intensity, gyroscope, magnetometer)
26
Data Stream © CSE AI Faculty Courtesy of G. Borriello
27
Activity Recognition Model © CSE AI Faculty Boosted cassifiers [Lester-Choudhury-etAl: IJCAI-05] Virtual evidence boosting [Liao-Choudhury-Fox-Kautz: IJCAI-07] Accuracy: 88% activities, 93% environment
28
Boosting Extremely flexible framework Handles high-dimensional continuous data Easy to implement Limitation: Only models local classification problems 28 © CSE AI Faculty
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.