Thread Subject:

unexpected biased in word recognition

I am going to use "traingda" to train a network for word recognition.
I was asked to recognition the image of six-letter words which is in the form of abccba, abbbba, aeggea......
and the images are divided into left and right half. PCA will be used for dimensional reduction on the left and right separately. Finally the PCs of the left and the right will be combined to the training set.

Then I need to test the images with only left or the right, while a blank images will be used instead of the original other half, e.g., abcxxx, xxxcba, abbxxx, xxxbba, (xxx represented to the blank images). I call these left (or the right) damaged images.

because the model are trained on the whole image, I expect that when classifying both left or right damaged images, the model gets only half of the images, so the accuracy will be lower than classifying the whole images. And the accuracies for classifying the left and right damage images should be more or less the same, cause they both have only half of the information from the image.

But the result is not like that. I found that the right-damage result are always 10% to 30 % better to the left-damage one.
I can't found any bugs in my own codes, and the function to generate the Principal component is given by my "boss", it should be correct. So I just can't explain why the bias is here.
I will be very appreciate if anyone can give me some advice.