Yes you could use a threshold for this. It is classifier specific - in this case we're using the KNN classifier, so you could perhaps use a distance threshold to place something into unknown.
Also, sometimes the right answer is not clear. If the number 6 is misclassified as 0, is that wrong? After all 6 is something that is like zero plus some extra stuff added on. It also sort-of resembles a 1. --Subutai On Sat, Sep 26, 2015 at 8:48 PM, CyanM <[email protected]> wrote: > Hello,Subutai > Thank you for you answer, I get it. > And these days, I test mnist_experiment, I got another question： > For example: my train set only contain handwritten digit ‘0’ and ‘1’, > but test set is ‘6', the classifier still classify the test image into ‘0’ > or ‘1’. > which means whatever the test image is, it will be classified into the > train set category? in other ward, the classifier will classify the test > image to the most similar category? > I think maybe there should be a threshold to control classifier to the > unknown category, am I right? > Thank you > Regard. > > Cyan >
