Schätzung der Konfidenz der Ausgabe eines Generativen Maschinenlernmodells

Details

Veröffentlichungs-Nr.

EP4797156

Anmeldetag

21. Februar 2025

Veröffentlichung

26. August 2026

Rechtsraum

EP

Abstract

A computer-implemented method (100) for estimating a level of confidence (4) of a response (3) that an artificial intelligence/machine learning model, Al/ML model (2), delivers in response to a given query (1), comprising the steps of: • obtaining (110) a representation (5) of the query (1) in a working space (6); • determining (120) a closeness (5a) of this representation (5) to one or more reference representations (8) of reference inputs (7) that are known to belong to a domain of validity, DOV (9), of the AI/ML model (2), and/or to this DOV (9) as a whole; • determining (130) the level of confidence (4) based at least in part on this closeness (5a); • and in response to the level of confidence (4) being (160) lower than a predetermined threshold: ∘ obtaining (170) feedback (3a) whether the response (3) provided by the Al/ML model (2) is correct; and ∘ if this feedback (3a) indicates (180) that the response (3) is correct, determining (190) that the representation (5) of the query (3) belongs to the DOV (9), wherein the query (1) relates (105) to at least one property of at least one industrial asset (11) in an industrial plant (10), and/or of an industrial process executed on this industrial plant (10).

Anmelder

Firma

ABB SCHWEIZ AG

Land

🇨🇭 Schweiz