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