分散知能を支えるネットワーキング

AIの大規模化に伴い，分散知能（Distributed Intelligence）が注目されています．特に，モデル分割に基づく推論や学習は，複数のノード群をネットワークを介して協調させることで実現される． 特に既存のMulti-hop split inference/ learning (MSI/MSL) では静的な通信経路に依存しているため，学習や推論のコンテキストに応じた柔軟なEnd-to-Endの通信経路の確立が求められています． 本研究では，Service Function Chaining (SFC) とMSIの類似性に着目し，MSI/MSLを支えるSFC基盤を確立することを目的としています¹． また，非同期的学習やMulti-path TCPを利用した高速な推論基盤の実現やモデル分割・配置・データルーチングの最適化問題の定式化についても検討しています．

Networking for Distributed Intelligence

With the increasing scale of AI, distributed intelligence has gained significant attention. In particular, inference and learning based on model splitting are realized by coordinating multiple nodes over a network. Existing Multi-hop split inference/learning (MSI/MSL) relies on static communication paths, necessitating the establishment of flexible end-to-end communication paths tailored to the context of learning and inference. Leveraging the similarity between Service Function Chaining (SFC) and MSI, this research aims to establish an SFC infrastructure for MSI/MSL¹. Additionally, we explore the realization of a high-speed inference/learning infrastructure using asynchronous learning and Multi-path TCP, as well as the formulation of optimization problems for model splitting, placement, and data routing.