Learning the moments data using a Cascade Architecture

The regionAttributes (feature vectors) generated in the previous experiment (Layer1_0_0_card_resize_p25-regionAttributes.mat) and the nearestSynapseDistance variable (Layer1_0_0_card_resize_p25-nearestSynapseDistance.mat) are used in this experiment

The variable sizes are 87492 x 595 (N xD format) and 87492 x 1 (N x1 format). The nearestSynapseDistance variable is converted to a boolean Y but using a inequality <=15 pixel distance. When the Decision tree based boosted classifiers are used as the individual nodes of the cascade classifier the results are as follows. The True positive rate should always be above 0.98 value but some how it keeps dipping below that minimum true positive rate set in the algorithm. The dip in the true positives in node 5 is also unexplained.

The next step is to train a Cascade classifier with the nodes based on Linear classifier.
QuadProg Freeware: The hex.sci.utah.edu's MATLAB installation doesn't have the optimization toolbox. So the normal quadprog () function is not available. The quadprog2 from Central file exchange has been used as the alternate solver. The solver can be found in the following location
http://www.mathworks.com/matlabcentral/fileexchange/loadFile.do?objectId=7860&objectType=FILE.
The solver takes H, X, A and b as the inputs. The solver takes lot of time for this dataset sometimes, and other times it quits telling that it is not able to find the optimum function value.