Memory Models for Incremental Learning Architectures. Viktor Losing, Heiko Wersing and Barbara Hammer

Transcription

1 Memory Models for Incremental Learning Architectures Viktor Losing, Heiko Wersing and Barbara Hammer

2 Outline Motivation Case study: Personalized Maneuver Prediction at Intersections Handling of Heterogeneous Concept Drift

3 Motivation Personalization adaptation to user habits / environments Lifelong-learning

4 Challenges - Personalized online learning Learning from few data

5 Challenges - Personalized online learning Learning from few data Sequential data with predefined order

6 Challenges - Personalized online learning Learning from few data Sequential data with predefined order Concept drift

7 Challenges - Personalized online learning Learning from few data Sequential data with predefined order Concept drift Cooperation between average and personalized model

8 Change is everywhere Coping with arbitrary changes

9 Change of taste / interest

10 Seasonal changes

11 Change of context

12 Rialto task: Change of lighting conditions

13 Setting Supervised stream classification Predict for an incoming stream of features x 1,, x j, x i R n the corresponding labels y 1, y j, y i {1,, c} On-line learning scheme After each touple x i, y i generate a new model h i to predict the next incoming example

14 Setting Supervised stream classification Predict for an incoming stream of features x 1,, x j, x i R n the corresponding labels y 1, y j, y i {1,, c} On-line learning scheme After each touple x i, y i generate a new model h i to predict the next incoming example

15 Setting Supervised stream classification Predict for an incoming stream of features x 1,, x j, x i R n the corresponding labels y 1, y j, y i {1,, c} On-line learning scheme After each touple x i, y i generate a new model h i to predict the next incoming example

16 Setting Supervised stream classification Predict for an incoming stream of features x 1,, x j, x i R n the corresponding labels y 1, y j, y i {1,, c} On-line learning scheme After each touple x i, y i generate a new model h i to predict the next incoming example

17 Setting Supervised stream classification Predict for an incoming stream of features x 1,, x j, x i R n the corresponding labels y 1, y j, y i {1,, c} On-line learning scheme After each touple x i, y i generate a new model h i to predict the next incoming example

18 Setting Supervised stream classification Predict for an incoming stream of features x 1,, x j, x i R n the corresponding labels y 1, y j, y i {1,, c} On-line learning scheme Preconditions for application: After each touple x i, y i generate a new model h i to predict the next incoming Obtainable example labels in retrospective

19 Definition Concept drift is given when the joint distribution changes t 0, t 1 : P t0 X, Y P t1 X, Y

20 Definition Concept drift is given when the joint distribution changes t 0, t 1 : P t0 X, Y P t1 X, Y t 0 Image source : Gama et al. A survey on concept drift adaptation, ACM Computing Surveys 2014

21 Definition Concept drift is given when the joint distribution changes t 0, t 1 : P t0 X, Y P t1 X, Y t 0 Real drift t 1 P Y X changes Image source : Gama et al. A survey on concept drift adaptation, ACM Computing Surveys 2014

22 Definition Concept drift is given when the joint distribution changes t 0, t 1 : P t0 X, Y P t1 X, Y t 0 Real drift t 1 Virtual drift P Y X changes P(X) changes Image source : Gama et al. A survey on concept drift adaptation, ACM Computing Surveys 2014

23 Definition Concept drift is given when the joint distribution changes t 0, t 1 : P t0 X, Y P t1 X, Y t 0 Real drift t 1 Virtual drift P Y X changes P(X) changes Image source : Gama et al. A survey on concept drift adaptation, ACM Computing Surveys 2014

24 Related work Dynamic sliding windows techniques PAW Bifet et al. Efficient Data Stream Classification Via Probabilistic Adaptive Windows, ACM 2013 Ensemble methods with various weighting schemes LVGB Bifet et al. Leveraging Bagging for Evolving Data Streams, ECML-PKDD 2010 Learn++.NSE Elwell et al. Incremental Learning in Non-Stationary Environments, IEEE-TNN 2011 DACC Jaber et al. Online Learning: Searching for the Best Forgetting Strategy Under Concept Drift, ICONIP-2013

25 Related work Dynamic sliding windows techniques PAW Bifet et al. Efficient Data Stream Classification Via Probabilistic Adaptive Windows, ACM 2013 Ensemble methods with various weighting schemes LVGB Bifet et al. Leveraging Bagging for Evolving Data Streams, ECML-PKDD 2010 Learn++.NSE Elwell et al. Incremental Learning in Non-Stationary Environments, IEEE-TNN 2011 DACC Jaber et al. Online Learning: Searching for the Best Forgetting Strategy Under Concept Drift, ICONIP-2013 Drawbacks: Target specific drift types Require hyperparameter setting according to the expected drift Discard former knowledge that still may be valuable

26 Drawbacks

27 Drawbacks Usual result

28 Drawbacks Desired behavior

29 Drawbacks Desired behavior

30 Drawbacks Desired behavior

31 Self Adaptive Memory (SAM)

32 Self Adaptive Memory (SAM) knn model knn model

33 Self Adaptive Memory (SAM)

34 Moving squares dataset

35 STM size adaptation

36 STM size adaptation Error %

37 STM size adaptation Error % %

38 STM size adaptation Error % % 7.12 %

39 STM size adaptation Error % % 7.12 % 0.0 %

40 STM size adaptation Error % % 7.12 % 0.0 %

41 Distance-based cleaning

42 Distance-based cleaning cleaning STM-consistent data

43 Distance-based cleaning STM Data to clean

44 Distance-based cleaning STM Data to clean

45 Distance-based cleaning STM Data to clean

46 Distance-based cleaning STM Data to clean

47 Distance-based cleaning STM Data to clean

48 Distance-based cleaning STM Data to clean

49 Adaptive compression cleaning STM-consistent data Long Term Memory

50 Adaptive compression cleaning STM-consistent data Long Term Memory class-wise clustering

51 Prediction

52 Prediction

53 Prediction

54 Prediction

55 Moving squares by SAM

56 Results: Error rates / ranks

57 SAM achieves best results

58 SAM is robust

59 Reasons for robustness Adaptation guided through error minimization Dynamic size of the STM Model selection for prediction Reduction of hyperparameters Consistency between STM and LTM LTM acts as safety net

60 Q & A