Verfahren zum Trainieren und Verwenden eines Neuronalen Netzwerks auf einer Edge-Vorrichtung durch Selektive Anpassung von Gewichten Gemäss Ihres Beitrags

Details

Veröffentlichungs-Nr.

EP4718329

Anmeldetag

30. September 2024

Veröffentlichung

1. April 2026

Rechtsraum

EP

Abstract

Methods are provided for training and using a neural network on an edge device of a communications network to perform classification. The neural network comprises layers and a plurality of binary or ternary weights, which are iteratively trained on a training data set comprising input samples and corresponding ground truth labels. The training comprises, during a forward pass, constructing contribution tensors being indicative of contributions of the weights of a layer towards an output of the layer; computing an output error of the neural network by comparing the output of a final layer to a ground truth label for an input sample; and during a backward pass, based on layer outputs obtained during the forward pass and the contribution tensors, identifying a set of weights which, when adapted, reduce the output error of the neural network, and adapting weights in that set to reduce the output error.

Anmelder

Firma

Robert Bosch GmbH

Land

🇩🇪 Deutschland