Activity Recognition from Accelerometers in Smart Homes

1. Activity Recognition from Accelerometers in Smart Homes Tom Diethe Manager, Core Machine Learning Amazon Research Cambridge Honorary Research Fellow Intelligent Systems Laboratory University of Bristol Bristol Robotics Lab, 16th August 2017

2. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References About Me 1996-2000 B.Sc. Experimental Psychology @ UoB 2000-2005 QinetiQ Ltd 2005-2006 M.Sc. Intelligent Systems @ UCL 2006-2010 Ph.D. Machine Learning & Signal Processing 2010-2012 Post-doc x2 @ UCL (CS + Stats) 2012-2013 Consultant @ BMJ Group 2013-2014 Researcher @ Microsoft Research Cambridge 2014-2017 Research Fellow @ UoB (SPHERE) Next stop: Amazon Research Cambridge Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

3. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References The SPHERE project Environmental temp, light, humidity, air quality, water & electricity Video emotion, gait, activity, interaction Wearable activity, sleep, etc. Contextual information demographics, medical history diaries, annotations Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

4. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References The SPHERE project Environmental temp, light, humidity, air quality, water & electricity Video emotion, gait, activity, interaction Wearable activity, sleep, etc. Contextual information demographics, medical history diaries, annotations Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

5. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Goal: Activity Recognition in Smart Homes Noisy data −→ methods that characterise the noise Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

6. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Goal: Activity Recognition in Smart Homes Noisy data −→ methods that characterise the noise Missing packets −→ robust methods Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

7. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Goal: Activity Recognition in Smart Homes Noisy data −→ methods that characterise the noise Missing packets −→ robust methods Limited storage and transmission capacity −→ sparse/compressive methods Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

8. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Goal: Activity Recognition in Smart Homes Noisy data −→ methods that characterise the noise Missing packets −→ robust methods Limited storage and transmission capacity −→ sparse/compressive methods Labelled data time-consuming and costly to acquire Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

9. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Goal: Activity Recognition in Smart Homes Noisy data −→ methods that characterise the noise Missing packets −→ robust methods Limited storage and transmission capacity −→ sparse/compressive methods Labelled data time-consuming and costly to acquire −→ Active Learning Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

10. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Goal: Activity Recognition in Smart Homes Noisy data −→ methods that characterise the noise Missing packets −→ robust methods Limited storage and transmission capacity −→ sparse/compressive methods Labelled data time-consuming and costly to acquire −→ Active Learning Diﬀering training and deployment settings Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

11. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Goal: Activity Recognition in Smart Homes Noisy data −→ methods that characterise the noise Missing packets −→ robust methods Limited storage and transmission capacity −→ sparse/compressive methods Labelled data time-consuming and costly to acquire −→ Active Learning Diﬀering training and deployment settings −→ Transfer Learning Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

12. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Motivation Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

13. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Motivation Would like a sparse representation of the data Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

14. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Motivation Would like a sparse representation of the data Don’t want to make parametric assumptions Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

15. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Motivation Would like a sparse representation of the data Don’t want to make parametric assumptions Features for classiﬁcation (activity recognition) Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

16. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Motivation Would like a sparse representation of the data Don’t want to make parametric assumptions Features for classiﬁcation (activity recognition) −→ Dictionary Learning Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

17. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Motivation Would like a sparse representation of the data Don’t want to make parametric assumptions Features for classiﬁcation (activity recognition) −→ Dictionary Learning Would like to incorporate with other models Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

18. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Motivation Would like a sparse representation of the data Don’t want to make parametric assumptions Features for classiﬁcation (activity recognition) −→ Dictionary Learning Would like to incorporate with other models −→ Bayesian approach Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

19. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Dictionary Learning Aim: ﬁnd a set of vectors di , (dictionary), to represent x ∈ Rn as a linear combination of these vectors: x = k i=1 zi di s.t. k n. (1) Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

20. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Dictionary Learning Aim: ﬁnd a set of vectors di , (dictionary), to represent x ∈ Rn as a linear combination of these vectors: x = k i=1 zi di s.t. k n. (1) Sparsity: few non-zero components zi (or many close to 0) Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

21. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Dictionary Learning Aim: ﬁnd a set of vectors di , (dictionary), to represent x ∈ Rn as a linear combination of these vectors: x = k i=1 zi di s.t. k n. (1) Sparsity: few non-zero components zi (or many close to 0) Cost function for m input vectors arranged in columns of matrix X ∈ Rn×m min Z,D X − DZ 2 F + λ n i=1 Ω(zi ) s.t. di 2 ≤ C, ∀i = 1, . . . , k. where D ∈ Rn×k is the dictionary, Z ∈ Rk×m are the coeﬃcients, and Ω(·) is a sparsity inducing regulariser, and λ determines the relative importance of good reconstructionsTom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

22. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Generative model for Equation (1) x N dot d z 0 0 ⌧ a b N N Ga(1, 1) Ga n m k Based on Yang et al. (2015). Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

23. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Generative model for Equation (1) x N dot d z 0 0 ⌧ a b N N Ga(1, 1) Ga Ga(1, 1) n m k Based on Yang et al. (2015). Gamma prior over β, which allows the dictionary atoms to be automatically scaled Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

24. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Generative model for Equation (1) x N dot d z ↵ 0 ⌧ a b N N Ga(1, 1) Ga N(0, 1) Ga(1, 1) n m k Based on Yang et al. (2015). Gamma prior over β, which allows the dictionary atoms to be automatically scaled additional level of hierarchy through the variables α on the means of the dictionary components, which can aid online learning Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

25. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Generative model for Equation (1) x N dot d z ↵ 0 ⌧ a b N N Ga(1, 1) Ga N(0, 1) Ga(1, 1) n m k Based on Yang et al. (2015). Gamma prior over β, which allows the dictionary atoms to be automatically scaled additional level of hierarchy through the variables α on the means of the dictionary components, which can aid online learning a, b control sparsity: Ga(1, 1) → non-sparse Ga(0.5, 10− 6) → sparse Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

26. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Infer.NET Mode (.NET code) Infer.NET Compiler Inference code Answers!Data Framework for running Bayesian inference in graphical modelsa Experiments used Monob, running on OS-X and Linux Inference engine: Variational Message Passing (VMP), deterministic approximation algorithm a http://research.microsoft.com/infernet b http://www.mono-project.com/ Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

27. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Modelling code p u b l i c void BDL( double [ , ] x , i n t numBases , double a , double b ) { var b a s i s = new Range ( numBases ) ; var s i g n a l = new Range ( x . GetLength ( 0 ) ) ; var sample = new Range ( x . GetLength ( 1 ) ) ; var n o i s e P r e c i s i o n = V a r i a b l e . GammaFromShapeAndRate (1 , 1 ) ; var c o e f f i c i e n t P r e c i s i o n s = V a r i a b l e . Array<double >(s i g n a l , b a s i s ) ; var dictionaryMeans = V a r i a b l e . Array<double >(b a s i s , sample ) ; var d i c t i o n a r y P r e c i s i o n s = V a r i a b l e . Array<double >(b a s i s , sample ) ; var c o e f f i c i e n t s = V a r i a b l e . Array<double >(s i g n a l , b a s i s ) ; var d i c t i o n a r y = V a r i a b l e . Array<double >(b a s i s , sample ) ; var s i g n a l s = V a r i a b l e . Array<double >(s i g n a l , sample ) ; dictionaryMeans [ b a s i s ] [ sample ] = V a r i a b l e . GaussianFromMeanAndVariance (0 , 1 ) . ForEach ( b a d i c t i o n a r y P r e c i s i o n s [ b a s i s , sample ] = V a r i a b l e . GammaFromShapeAndRate (1 , 1 ) . ForEach ( b a s c o e f f i c i e n t P r e c i s i o n s [ s i g n a l , b a s i s ] = V a r i a b l e . GammaFromShapeAndRate ( a , b ) . ForEach ( s i c o e f f i c i e n t s [ s i g n a l , b a s i s ] = V a r i a b l e . GaussianFromMeanAndPrecision (0 , c o e f f i c i e n t P r e c d i c t i o n a r y [ b a s i s , sample ] = V a r i a b l e . GaussianFromMeanAndPrecision ( dictionaryMeans [ b a s i s , sample ] , d i c t i o n a r y P r e c i s i o n s [ b a s i s , sample ] ) ; d i c t i o n a r y [ b a s i s , sample ] . I n i t i a l i s e T o ( d i c t i o n a r y P r i o r s [ b a s i s , sample ] ) ; var c l e a n S i g n a l s = V a r i a b l e . M a t r i x M u l t i p l y ( c o e f f i c i e n t s , d i c t i o n a r y ) . Named( ” c l e a n ” ) ; s i g n a l s [ s i g n a l , sample ] = V a r i a b l e . GaussianFromMeanAndPrecision ( c l e a n S i g n a l s [ s i g n a l , s s i g n a l s . ObservedValue = x ; } Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

28. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Datasets HAR dataset Anguita et al. (2013) Publicly available activity recognition dataset1 Smart-phone on the waist, with tri-axial accelerometer, 50 Hz Annotation was done using video-recordings. Activities: Lie, Stand, Walk, Sit, Ascend stairs, Descend stairs SPHERE challenge dataset Twomey et al. (2016) collected by SPHERE project and made public as a challenge2 Tri-axial accelerometer on dominant wrist, 20 Hz, range ±8 g Activites: Lie, Stand, Walk 1 https://archive.ics.uci.edu/ml/datasets/Human+Activity+ Recognition+Using+Smartphones 2 http://irc-sphere.ac.uk/sphere-challenge/home Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

29. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Sparsity Non-sparse priors (Ga(1, 1)) average sparsity 0.84 bases signals Sparse priors (Ga(0.5, 10−6)) average sparsity 0.96 bases signals Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

30. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Convergence of VMP Model evidence converges monotonically to a local maximum 0 2 4 6 8 10 12 14 16 18 Iterations 350000 300000 250000 200000 150000 100000 50000 0 50000 Evidencelogodds non-sparse sparse Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

31. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Convergence of reconstruction error 0 2 4 6 8 10 12 14 16 18 Iterations 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 RMSE non-sparse sparse Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

32. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Learnt Dictionary 150000 100000 50000 0 50000 100000 150000 150000 100000 50000 0 50000 100000 150000 150000 100000 50000 0 50000 100000 150000 20020406080100120140160 150000 100000 50000 0 50000 100000 150000 20020406080100120140160 20020406080100120140160 20020406080100120140160 Dictionary 0-16 x y Figure: 16 example bases from the dictionary of 128 bases inferred by BDL using the non-sparse priors on HAR dataset Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

33. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Coeﬃcients Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

34. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Reconstructions 0.6 0.4 0.2 0.0 0.2 0.4 0.6 0.8 0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.6 0.4 0.2 0.0 0.2 0.4 0.6 0.8 signal reconstruction ±s.d. Reconstructions, RMSE=0.0276 time (s) magnitude Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

35. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Reconstruction Performance SPAMS BDL Sparse BDL Bases RMSE Sparsity RMSE Sparsity RMSE Sparsity 64 0.0480 0.71 0.0519 0.88 0.0293 0.62 128 0.0457 0.84 0.0400 0.94 0.0276 0.84 256 0.0438 0.92 0.0316 0.99 0.0288 0.93 512 0.0423 0.96 0.0224 0.99 0.0231 0.96 Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

36. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Classiﬁcation Multi-class Bayes Point Machine (BPM) Herbrich et al. (2001), also using Infer.NET Assumptions: Feature values x are always fully observed The order of instances does not matter The predictive distribution is a linear discriminant of the form: p(yi |xi , w) = p(yi |si = w xi ), where w are the weights and si is the score for instance i y ˜s s 1 N w x ⇥ µw ⌧w N N D A R Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

37. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Activity Recognition Results Use coeﬃcients as features of classiﬁcation algorithm 64 bases + bias feature Metric: per-class 1 vs rest area under ROC curve HAR dataset Activity BDL sparse BDL SPAMS Walking 0.73 0.83 0.88 Ascending stairs 0.63 0.60 0.83 Descending stairs 0.61 0.34 0.82 Sitting 0.74 0.72 0.89 Standing 0.51 0.43 0.98 Lying down 0.95 0.95 0.95 Average 0.70 0.65 0.89 SPHERE dataset Activity BDL sparse BDL SPAMS Walking 0.55 0.51 0.46 Standing 0.69 0.62 0.44 Lying down 0.86 0.85 0.46 Average 0.70 0.66 0.45 Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

38. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Transfer Learning Source: DS = {(xi , yi )mS i=1} Target: DT = {(xi , yi )mT i=1} Assumptions: labels available for source domain labels can be acquired for target domain, but costly Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

39. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Hierarchical Model Multiclass Bayes Point Machine (BPM) y ˜s s 1 N w x ⇥ µw ⌧w N N D A R Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

40. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Hierarchical Model Community: Learn community posteriors y ˜s s 1 N w x ⇥ µw ⌧w µµ µ k⌧ ✓⌧ N N N D A R Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

41. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Hierarchical Model Transfer: weight means and precisions replaced by community posteriors from source y ˜s s 1 N w x ⇥ µw ⌧w N N D A R Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

42. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Hierarchical Model Personalisation: smoothly evolve from generic to custom predictions y ˜s s 1 N w x ⇥ µw ⌧w N N D A R Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

43. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Active Learning Baseline: random (= online learning) Uncertainty Sampling Nearest to decision boundary, fails in case of corrupted data Certainty Sampling Furthest from decision boundary. Normally nonsensical but solves corruption issue Value of Information (VOI) Kapoor et al. (2007) Decision theoretic measure that uses p(y = 1) vs p(y = 0) (relative risk). Expensive to compute. Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

44. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Datasets Source: Anguita et al. (2013) 30 subjects, Smartphone, 50Hz, Video annotations Target: Zhang and Sawchuk (2012) 14 subjects, MotionNode, 100Hz, Observer annotations Classes: Walking upstairs vs. Walking downstairs Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

45. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Results 0 5 10 15 20 # instances 0.0 0.2 0.4 0.6 0.8 1.0Accuracy Hold-out Accuracy Random US CS VOI+ VOI- Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

46. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Conclusions and Further Work Improved methods for Bayesian Dictionary Learning Sparsity does not need to be enforced Application to accelerometer signals for Activity Recognition Hierarchical extension to BPM for transfer learning Combination of Active & Transfer learning −→ fast learning on target dataset Uncertainty Sampling good enough Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

47. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Conclusions and Further Work Improved methods for Bayesian Dictionary Learning Sparsity does not need to be enforced Application to accelerometer signals for Activity Recognition Hierarchical extension to BPM for transfer learning Combination of Active & Transfer learning −→ fast learning on target dataset Uncertainty Sampling good enough Further Work: Incorporate classiﬁer directly into the model Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

48. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Conclusions and Further Work Improved methods for Bayesian Dictionary Learning Sparsity does not need to be enforced Application to accelerometer signals for Activity Recognition Hierarchical extension to BPM for transfer learning Combination of Active & Transfer learning −→ fast learning on target dataset Uncertainty Sampling good enough Further Work: Incorporate classiﬁer directly into the model Source code: https://github.com/IRC-SPHERE/bayesian-dictionary-learning https://github.com/IRC-SPHERE/ActiveTransfer Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes

49. Introduction Methods Bayesian Dictionary Learning Infer.NET implementation Experiments Conclusions References Selected References D Anguita, A Ghio, L Oneto, X Parra, and JL Reyes-Ortiz. A public domain dataset for human activity recognition using smartphones. In ESANN, 2013. Tom Diethe, Niall Twomey, and Peter Flach. Active transfer learning for activity recognition. In 24th European Symposium on Artiﬁcial Neural Networks, ESANN 2016, Bruges, Belgium, April 27-29, 2016, 2016a. Tom Diethe, Niall Twomey, and Peter Flach. BDL. NET: Bayesian dictionary learning in Infer. NET. In Machine Learning for Signal Processing (MLSP), 2016 IEEE 26th International Workshop on, pages 1–6. IEEE, 2016b. Ralf Herbrich, Thore Graepel, and Colin Campbell. Bayes point machines. Journal of Machine Learning Research, 1:245–279, January 2001. URL http://research.microsoft.com/apps/pubs/default.aspx?id=65611. Ashish Kapoor, Eric Horvitz, and Sumit Basu. Selective supervision: Guiding supervised learning with decision-theoretic active learning. In IJCAI, volume 7, pages 877–882, 2007. Niall Twomey, Tom Diethe, Meelis Kull, Hao Song, Massimo Camplani, Sion Hannuna, Xenofon Fafoutis, Ni Zhu, Pete Woznowski, Peter Flach, and Ian Craddock. The sphere challenge: Activity recognition with multimodal sensor data. arXiv preprint arXiv:1603.00797, 2016. Linxiao Yang, Jun Fang, Hong Cheng, and Hongbin Li. Sparse Bayesian dictionary learning with a Gaussian hierarchical model. CoRR, abs/1503.02144, 2015. URL http://arxiv.org/abs/1503.02144. M. Zhang and A.A. Sawchuk. USC-HAD: A daily activity dataset for ubiquitous activity recognition using wearable sensors. In ACM Int. Conf. on Ubiquitous Computing Workshop on Situation, Activity and Goal Awareness (SAGAware), 2012. Tom Diethe Amazon Research Cambridge; University of Bristol Activity Recognition from Accelerometers in Smart Homes