/ Electro-Optical Communication (ECO) group
Contact: Decebal Constantin Mocanu
e-mail: d.c.mocanu@tue.nl
The Challenge
Factored four-way conditional restricted Boltzmann
machines (FFW-CRBMs) for activity recognition
D. C. Mocanu
1
, H. Bou Ammar
2
, D. Lowet
3
, K. Driessens
4
, A. Liotta
1
, G. Weiss
4
, K. Tuyls
5
1
Eindhoven University of Technology, Netherlands
2
University of Pennsylvania, USA
3
Philips Research, Eindhoven, Netherlands
4
Maastricht University, Netherlands
5
University of Liverpool, United Kingdom
SNN Symposium
Intelligent Machines 2015
Experiments and Results
Conclusion
• Novel machine learning technique
for
activity
recognition
and
prediction suitable for any time
series in general.
• FFW-CRBM came together with an
adapted training algorithm SMcCD.
• FFW-CRBMs are capable of: : (1)
classification, (2) prediction, and
(3) self auto evaluation of their
classification performance within
one unified framework.
The Solution
The Exploitation
Example on Fibonacci Series
Reference
[1] D. C. Mocanu, H. B. Ammar, D. Lowet, K. Driessens, A. Liotta, G.
Weiss, and K. Tuyls, “Factored four way conditional restricted Boltzmann
machines for activity recognition,” Pattern Recognition Letters, 2015.
Method Activities EA1 EA2 EA3 EA4 EA5 EA6
FFW-CRBM EA1 98 ± 1.5 1.3 ± 0.8 0 0 0 0.7 ± 0.6 EA2 1.3 ± 1.2 98.7 ± 1.2 0 0 0 0 EA3 0.6 ± 0.9 0 94 ± 2.6 1.2 ± 1.1 4.2 ± 0.7 0 EA4 6 ± 1.2 0 0 94 ± 1.2 0 0 EA5 0 0 0 0 98.9 ± 1.1 1.1 ± 1.1 EA6 0 0 7.4 ± 2.1 0 3.6 ± 1.6 89 ± 2.3 SVM EA1 94 ± 2.1 0 0 0 0.7 ± 0.8 5.3 ± 1.9 EA2 8.5 ± 3.1 91.5 ± 3.1 0 0 0 0 EA3 15 ± 2.3 2.3 ± 1.4 82.1 ± 3.4 0 0.6 ± 0.8 0 EA4 47 ± 5.3 1 ± 1.2 0 52 ± 4.7 0 0 EA5 6 ± 1.6 0 0 0 94 ± 1.6 0 EA6 28 ± 2.3 3.1 ± 1.6 0 1.9 ± 1.4 2 ± 1.8 65 ± 2.7 Activities CRBM FCRBM FFW-CRBM EA1 0.110 ± 0.005 0.054 ± 0.002 0.028 ± 0.008 EA2 0.138 ± 0.003 0.036 ± 0.006 0.018 ± 0.012 EA3 0.106 ± 0.012 0.044 ± 0.021 0.023 ± 0.011 EA4 0.126 ± 0.011 0.094 ± 0.008 0.027 ± 0.014 EA5 0.125 ± 0.004 0.068 ± 0.011 0.026 ± 0.009 EA6 0.123 ± 0.026 0.093 ± 0.007 0.048 ± 0.019
I. Exercise activities
“body squats” (EA1), “vertical
to horizontal hand movements”
(EA2), “opening and closing of
arms while moving” (EA3),
“jumping” (EA4), “leg lunges”
(EA5), and “walking” (EA6)
a. Classification accuracy [%]
b. Prediction - RMSE
One-step prediction
c. Self auto evaluation
CRBM FCRBM FFW-CRBM 10 steps 0.116 ± 0.007 0.048 ± 0.016 0.038 ± 0.021 20 steps 0.121 ± 0.008 0.046 ± 0.019 0.043 ± 0.026 30 steps 0.121 ± 0.049 0.049 ± 0.021 0.047 ± 0.017 40 steps 0.118 ± 0.011 0.059 ± 0.035 0.056 ± 0.028 50 steps 0.119 ± 0.012 0.086 ± 0.102 0.059 ± 0.027