Skip to main content
8 events
when toggle format what by license comment
Jan 31, 2020 at 15:14 comment added mrkwse An approach may be to use/train some form of binary classifier to discard non-digit examples prior to the recognition step (i.e., 0-9 classification). You'll have to maximise recall of the digit inputs, so there may still be a few non-digits slip through, but such an approach may reduce the detriment to the performance of classifying the real digits.
Jan 30, 2020 at 9:30 comment added MSalters Point from practical experience: having one other class may work, but in some cases it can help to have multiple. I'm not aware of a method to determine up front what the best number of "other" classes is.
Jan 30, 2020 at 0:17 comment added Apollys supports Monica Yes, this is an important point. Once again, the "other" class will have to come from some specific distribution and have enough samples to cover that distribution. That means if you really want your model to learn what it means to be a digit rather than anything else that is possible, you will need a (probably intractably) lot of data. In this way, we see that deep learning is largely about understanding and confining the problem domain very well.
Jan 29, 2020 at 20:50 comment added Alexander Soare Totally agree here. I was thinking something along the lines when I saw the suggestion. @MatiasValdenegro (don't know if tagging you here works), I was wondering what your thoughts on such an approach might be.
Jan 29, 2020 at 20:08 comment added Kyle A This answer looks like it should be a comment, but it looks like you don't have enough reputation to comment. You brought up some great concerns about the effectiveness of adding an "other" class. It can be really hard to predict what will or won't work well with neural networks.
Jan 29, 2020 at 14:30 review Low quality posts
Jan 31, 2020 at 0:20
Jan 29, 2020 at 14:10 review First posts
Jan 29, 2020 at 19:58
Jan 29, 2020 at 14:07 history answered Rich Chase CC BY-SA 4.0