For the binary classifier we can reduce output to a single number, the (log) prob of being something rather than nothing.
By thresholding this number with a varying threshold, we can generate a receiver operator curve to describe performance.
We can also make histograms for this number for the four bins of the confusion matrix.
Finally, we can investigate the images that lie on the boundary and possibly identify labellinig errors.
For the binary classifier we can reduce output to a single number, the (log) prob of being something rather than nothing.
By thresholding this number with a varying threshold, we can generate a receiver operator curve to describe performance.
We can also make histograms for this number for the four bins of the confusion matrix.
Finally, we can investigate the images that lie on the boundary and possibly identify labellinig errors.