Vortraining eines Codierers mit Unmarkierten Daten durch Maskierte Eigenschaftsvorhersage

Details

Veröffentlichungs-Nr.

EP4790591

Anmeldetag

5. Februar 2025

Veröffentlichung

12. August 2026

Rechtsraum

EP

Abstract

A method (100) for training an encoder (1) that is configured for encoding a record (2) of measurement data into a representation (3) in a working space, the measurement data comprising measurement values that are associated with positions in space, the method (100) comprising the steps of:• providing (110) at least one training record (2a) of measurement data;• designating (120) cells (4) in space, wherein at least some of the cells (4) contain positions in space with which measurement values in the training record (2a) of measurement data are associated;• determining (130), for each cell (4), based on the training record (2a) of measurement data, an indication (4a) of a predetermined property;• creating (140) a masked record (2#) of measurement data by removing, from the training record (2a) of measurement data, all measurement values that are associated with positions in space that lie in at least one to-be-masked cell (4#);• providing (150) the masked record (2#) of measurement data to the to-be-trained encoder (1), thereby obtaining a representation (3);• decoding (160), by a decoder (5), the representation (3) into a reconstructed indication (4a*) of the predetermined property for at most neighboring cells of not-masked cells (4);• rating (170), by means of a predetermined loss function (6), how well the reconstructed indication (4a*)of the predetermined property corresponds to the original indication (4a) of the predetermined property derived from the training record (2a) of measurement data; and• optimizing (180) parameters (1a) that characterize the behavior of the encoder (1) towards the goal of improving the rating (6a) by the loss function (6) upon further processing of training records (2a).

Anmelder

Firma

Robert Bosch GmbH

Land

🇩🇪 Deutschland