In the CalibrationAUC, the standard_devs are defined by:
standard_devs = [np.abs(set_['error'])/set_['confidence'] for set_ in data['sets_by_confidence']]
I am confused about the set_['confidence'] here because the confidence is calculated by 1. / ((alphas-1) * lambdas) as in the "predict.py" file, while this value is not the square root of the variance of mean values (betas / ((alphas-1) * lambdas).
By the way, I also noticed that the confidence calculated here is different from the uncertainty defined in Figure 2B of your paper, may I ask why using this metric (1. / ((alphas-1) * lambdas)) to evaluate confidence (or uncertainty).
In your repository of "evidential-deep-learning", I found the calibration plot is drawn with the standard deviation betas / ((alphas-1) * lambdas), and the confidence is also measured by this value. I wonder why it changes in these two repositories.
In the CalibrationAUC, the standard_devs are defined by:
standard_devs = [np.abs(set_['error'])/set_['confidence'] for set_ in data['sets_by_confidence']]I am confused about the
set_['confidence']here because the confidence is calculated by1. / ((alphas-1) * lambdas)as in the "predict.py" file, while this value is not the square root of the variance of mean values (betas / ((alphas-1) * lambdas).By the way, I also noticed that the confidence calculated here is different from the uncertainty defined in Figure 2B of your paper, may I ask why using this metric (
1. / ((alphas-1) * lambdas)) to evaluate confidence (or uncertainty).In your repository of "evidential-deep-learning", I found the calibration plot is drawn with the standard deviation
betas / ((alphas-1) * lambdas), and the confidence is also measured by this value. I wonder why it changes in these two repositories.