Föderiertes Kontinuierliches Lernverfahren für ein Grosses Multimodales Modell und Lernvorrichtung damit

Details

Veröffentlichungs-Nr.

EP4749534

Anmeldetag

16. Dezember 2024

Veröffentlichung

27. Mai 2026

Rechtsraum

EP

Abstract

The present disclosure provides a federated continual learning method for a large multimodal model, comprising steps of: inputting at least one training data to the m-th local large multimodal model to (i) generate an embedding vector, (ii) (ii_1) generate an (m_1)-th local prompt vector and an (m_1)-th global prompt vector, and merge the (m_1)-th local prompt vector and the (m_1)-th global prompt vector to thereby generate an (m_1)-th merged prompt vector, and (ii_2) generate a first hidden status vector, (iii) (iii_1) generate an (m_j)-th local prompt vector and an (m_j)-th global prompt vector, and merge the (m_j)-th local prompt vector and the (m_j)-th global prompt vector to thereby generate an (m_j)-th merged prompt vector, and (iii_2) generate a j-th hidden status vector, and (iv) repeat learning iteration to the training data; and transmitting updated information of the local prompt generators to thereby update the global prompt generators.

Anmelder

Firmen

Seoul National University R&DB Foundation
UIF (University Industry Foundation), Yonsei University

Land

🇰🇷 Südkorea