Speichereffizientes Maschinenlernmodelltraining

Details

Veröffentlichungs-Nr.

EP4760601

Anmeldetag

3. Dezember 2025

Veröffentlichung

17. Juni 2026

Rechtsraum

EP

Abstract

Various strategies for training a machine learning model are provided. One method computes a gradient matrix of a training loss function and applies a whitening transformation. Gradient normalization and whitening are employed to update weight vectors solely based on current gradient information without the need for exponential moving averages. A second method determines a state matrix to derive a truncated basis matrix composed of selected principal eigenvectors and additional orthogonal vectors and calculates an updated weight by projecting the gradient onto this low-dimensional subspace.

Anmelder

Firma

Microsoft Technology Licensing, LLC

Land

🇺🇸 USA