Difference between svmtrain in R2013a and R2014a

I have written a code and created a labeled and normalized database. I wanted to classify it by SVM using majority voting. I have used 2 types of SVM using svmtrain function, linear and RBF kernels. Note that I used the database for training. The problem is that when I run the same code in Matlab R2013a and R2014a in R2013a, the linear kernel give me 93% CCR (Correct Classification Ratio) but in R2014 it gives me 78% CCR. I checked the settings in both versions even the svmtrain source code, they are the same but cannot understand what is the reason of this big difference! I appreciate it if anyone could help me understand what is the difference. Maybe it is because of a bug in R2014a optimization process used in svmtrain? Attached you can find the code files and the database. Run the "OMR_TrainingClassifier.m"

.

program output in R2013a:
CCR for training Datas for SVM:
    0.9307

CCR for training Datas for RBF:
    0.9966

program output in R2014a:

CCR for training Datas for SVM:
    0.7759

CCR for training Datas for RBF:
    0.9966

[Ilya]

It is hard to isolate the problem because your script is complex and you train many classifiers. Post simple reproduction steps. Find one binary classifier that gives different predictions in 13a and 14a.