Gender recognition from frontal facial images within project FP7-ICT-247525-HUMAVIPS

Experiment 05 learned by parallel BMRM

<< zpět


In this experiment will be used extended version of BMRM that split dataset into N groups. From every group will be created cutting plane in every iteration. If we add more cutting planes algorithm will need linearly more memory. Aim of experiment is to compare ordinary BMRM with its extension.

For comparition it will be used only small part (3 000 samples) of whole dataset and only a few tranformations (for time-saving). In this experiment dataset will be splitted into four groups, so in every iteration will be added four cutting planes.

Results LOG from process

Parameters of experiment
imSize [h x w]80 x 64
winSize [h x w]60 x 40
window positionhalf in either axis
window rotation{-pi/30; 0; pi/30}
window scale{1}
LAMBDAOrdinary BMRMExtended BMRM
Training errorValidation errornIterationsTraining errorValidation errornIterationsUsed memory [MB]
2 1.33%11.46%35 1.40%11.46%351 547
1 0.03%9.98%58 0.03%9.93%602 651
0.5 0.00%9.80%130 0.00%9.70%1295 701
0.1 0.00%9.71%405 0.00%9.70%37616 627
Testing err.17.53% 17.35%

Figure 1: Criteria function ... also in MATLAB figure


Choon Hui Teo , S.V.N. Vishwanthan , Alex J. Smola , Quoc V. Le: Bundle Methods for Regularized Risk Minimization, The Journal of Machine Learning Research, 11, p.311-365, 3/1/2010
<< zpět