How do I know which data I predicted wrong and improve the model prediction accuracy

[zhe wang]

Fine-tuned a CNN using caffe, I got my model and the accuracy.

But how do I know which data I predicted is wrong and how could I improve the model or parameter by using these wrong-predicted data?

Thanks for answering

[zhe wang]

I thought for a night, maybe discarding the last loss layer and feed in data seperately, we can get the classification.... Now the problem arise, how do we know the classification is right

在 2014年10月9日星期四UTC+8上午12时15分07秒，zhe wang写道：

[lixin7...@gmail.com]

seems python wrapper could do this.

not s

在 2014年10月9日星期四UTC+8上午12时15分07秒，zhe wang写道：

[lixin7...@gmail.com]

not sure

在 2014年10月9日星期四UTC+8上午8时47分26秒，zhe wang写道：

[zhe wang]

Thanks for answering. I have compiled matlab interface and using matlab to solve it. Matlab interface can give each sample a score and I can compare it with my label to know which prediction is wrong. Thus feeding more of this kind of sample, I may get a better model.

在 2014年10月11日星期六UTC+8上午2时03分50秒，lixin7...@gmail.com写道：