DEUPDecomposition

class probly.method.deup.DEUPDecomposition(representation: DEUPRepresentation[T], base: LogBase = None)[source]

Bases: AdditiveDecomposition, Generic

DEUP additive uncertainty decomposition.

Follows the classification setup of [LJN+23], where aleatoric uncertainty \(H[P(Y|x)]\) of the ground-truth conditional is taken to be zero in the absence of label replicates:

  • total = predicted expected cross-entropy (output of the error head),

  • aleatoric = 0,

  • epistemic = total = predicted expected cross-entropy.

The additive constraint total = aleatoric + epistemic therefore holds by construction. Note that, contrary to a common shortcut, the entropy of the predictive softmax is not the aleatoric uncertainty in the DEUP definition: the paper defines aleatoric as the entropy of the (unknown) ground-truth conditional, not of the learner’s estimate.

property aleatoric: AU

The aleatoric uncertainty of the decomposition.

base: LogBase
canonical_notion[source]

alias of TotalUncertainty

property components: list[type[Notion]]

The components of the decomposition.

property epistemic: EU

The epistemic uncertainty of the decomposition.

get(k[, d]) → D[k] if k in D, else d.  d defaults to None.
get_canonical() → Notion[source]

Return the canonical notion of the decomposition.

get_notion(notion: type[N]) → N[source]

Return the component corresponding to the given notion.

items() → a set-like object providing a view on D's items
keys() → a set-like object providing a view on D's keys
representation: DEUPRepresentation[T]
property total: TU

The total uncertainty of the decomposition.

values() → an object providing a view on D's values

Examples using probly.method.deup.DEUPDecomposition

DEUP on MNIST

DEUP on MNIST