What question did this study set out to answer?

This deliverable investigates methods for efficiently managing machine learning models in future 6G wireless networks.

March 5, 20260 citationsOpen Access

Deliverable D3.3 - Model Caching, Training, and Compression

Q: What does this research mean for the field?

Efficient management of Machine Learning models through coordinated caching, training, and compression can enable scalable, low-latency, and energy-efficient AI services in Sixth Generation (6G) wireless networks. Novelty: ClaimNovelty.NOVEL_FINDING. Consensus alignment: ConsensusAlignment.NEUTRAL.

6P6G-GOALS project

Key Points

This deliverable investigates methods for efficiently managing machine learning models in future 6G wireless networks.
Explored model caching, training, and compression across five application frameworks.
Focused on context-aware data transmission to reduce traffic in IoT devices.
Developed neuro-inspired solutions for multimodal inference in audio-visual systems.
Introduced split-training methodologies for transformer architectures.
Proposed reinforcement learning methods and implicit neural representations for effective model compression.
Achieved reduced redundant data traffic through context-aware processes.
Enhanced temporal coherence in multimodal data despite delays.
Lowered communication overhead without sacrificing accuracy in split learning.
Balanced performance and resource costs while managing latencies.
Provided flexible, compressed representations of channel state information suitable for 6G.

Abstract

This deliverable investigates how future Sixth Generation (6G) wireless networks can efficiently manage Machine Learning (ML) models through coordinated (i) caching, (ii) training, and (iii) compression across the protocol stack, a key objective of the 6G‑GOALS project in realizing Artificial Intelligence (AI)‑native networks. The study spans five application frameworks, each targeting a specific subset of these model‑management functions. Contribution A focuses on model caching and model compression within the Semantic Data Acquisition (SEMDAS) framework, where a shared embedding space and semantic relevance scoring govern when distributed Internet‑of‑Things (IoT) devices transmit data, turning random access into a context‑aware, collision‑sensitive process that reduces redundant traffic. Contribution B targets model caching for real‑time multimodal inference in distributed Audio‑Visual (AV) systems, proposing a neuro‑inspired solution that maintains temporal coherence of multimodal received data under stochastic, modality‑specific delays while leveraging cached multimodal foundation models at the edge. Contribution C addresses model training and model compression by introducing an efficient split‑training methodology for transformer‑based architectures, combining batch‑ and token‑level compression to lower communication overhead in Split Learning (SL) without sacrificing accuracy. Contribution D advances model training and model compression through Reinforcement Learning methods for delayed and constrained systems, where decisions must balance performance and resource costs under non‑negligible observation and actuation latencies. Finally, Contribution E focuses on model compression by proposing implicit neural representations (INR)‑based scheme for Channel State Information (CSI) compression, integrating quantisation and entropy coding to provide flexible, highly compressed CSI representations suitable for 6G links. Together, these contributions show how model caching, training, and compression can be jointly orchestrated across diverse frameworks to enable scalable, low‑latency, and energy‑efficient AI services in goal‑oriented semantic communication networks.

问 AI

Bookmark

View Full Paper