Implemented methods¶
The following methods are currently implemented in probly.
Representation¶
Second-order Distributions¶
These methods represent (epistemic) uncertainty by a second-order distribution over distributions.
Bayesian Neural Networks¶
[BCKW15]
Dropout¶
[GG16b]
DropConnect¶
[MNM+19]
Deep Ensembles¶
[LPB17b]
Evidential Deep Learning¶
Credal Sets¶
These methods represent (epistemic) uncertainty by a convex set of distributions.
Credal Ensembling¶
[NZD25]
Quantification¶
Upper / lower entropy¶
Generalized Hartley¶
Entropy-based¶
Distance-based¶
Conformal Prediction¶
These methods represent uncertainty by a set of predictions.
Split Conformal Prediction¶
[AB21]
Calibration¶
These methods adjust the model’s probabilities to better reflect the true probabilities.
Focal Loss¶
[LGG+17]
Label Relaxation¶
Temperature Scaling¶
[GPSW17b]