1 Reply Latest reply on Dec 30, 2016 2:26 PM by Intel Corporation

    101 Curie NN questions about effectiveness - incorrect classification on the SAME training data

    marcuso

      Hi,

       

      I have processed some OCR figures (I will post a paper on this soon on my blog), the images are now contained in a <127 uint 8 array, so they can be trained into the currie 101 NN hardware NN.

       

      My question is this

       

      I train my NN with a set of data, it learns 10 images, and has 10 committed neurons

      I then test my NN with the same data, and it recognises 19 out of 20 images correctly, one it gets wrong.

       

      Is this normal/ok ?

       

      Also...

       

      Can I reduce this error by repeatedly training the same images say 10 times, so learn with the same data 10 times etc

      What is better KNN or RBF ?

      If it gets one wrong by classifing it incorrectly should I use category 0 to unlearn it, it say it gets image 16 wrong, the image is of a cat 7, when I ask the NN to classify the same image it comes back with a cat 8.

       

      Should I get this image that it classifies wrong, and learn it into cat 0 ? Will this mean that it no longer will give me an incorrect answer, is there a general strategy for this situation ?

       

      Thanks.

       

      Marcus