Effiziente Nutzung von Verarbeitungsressourcen für Maschinelles Lernen

Details

Veröffentlichungs-Nr.

EP4756675

Anmeldetag

1. Dezember 2025

Veröffentlichung

10. Juni 2026

Rechtsraum

EP

IPC

G06N3/0464, G06N3/063

Abstract

There are provided measures for efficient utilization of machine learning processing resources. Such measures exemplarily comprise, for preparing, based on initial data, input data for an input layer of a machine learning model having said input layer, at least one inner layer, and an output layer, and being learned and/or deployed at processing hardware providing a plurality of processing channels, each processing channel comprising a dedicated processor and a dedicated memory range, wherein said initial data is arranged in a first array having a first dimension equal to or larger than 1, a second dimension equal to or larger than 1, and a third dimension equal to or larger than 1, wherein a data amount of each element of said first array is equal to a predetermined amount, if said predetermined amount multiplied with a size of said first dimension and a size of said second dimension exceeds a size of said dedicated memory range and if said third dimension is smaller than a number of said plurality of processing channels: partitioning said first array in said first dimension and said second dimension into a plurality of uniform sub-arrays, wherein a number of said plurality of sub-arrays is based on said dedicated memory range, performing, for each sub-array field index position of a plurality of selected sub-array field index positions, a down-sampling operation on said first array of said initial data, said down-sampling operation includes selecting, from each sub-array, a field at said sub-array field index position, and generating, for each index position in said third dimension, a second array including said elements corresponding to said selected fields and said index position in said third dimension, in an arrangement corresponding to an arrangement of said plurality of uniform sub-arrays, and assigning each of generated second arrays in order of generation to a respective processing channel of said plurality of processing channels as said input data.

Anmelder

Firma

Nokia Solutions and Networks Oy

Land

🇫🇮 Finnland