I am looking into a numeral classification (Devanagari OCR). I am training data on PECL SVM model in php. To plot the learning curve I started testing robustness of SVM by 5 fold cross validation. I got result like below -
When less number of dataset is there it gives high accuracy but while dataset size increased accuracy decreased gradually. I know SVM works well in less number of dataset. But my project supervisor is not satisfied with that. I searched over internet then I found nothing about such problem. Is there any problem in my Model(SVM with RBF kernel provided by PECL SVM)? or is it Normal as property of SVM. Classifying new image by model is perfect.