OMNIS+: Semantic RAN Slicing and Quantization for Dynamic Split Neural Networks

Langtian Qin, Avishkar Dnyaneshwar Kawadgave, Ian Andrew Harshbarger, Leïla Nasraoui, Carla Fabiana Chiasserini, Marco LevoratoarXiv 2026 · arXiv

TL;DR

OMNIS+ is a novel semantic communication framework that optimizes edge-based machine learning by combining dynamic split neural networks with a resilient "box" quantization scheme. By jointly managing radio access network (RAN) slicing and neural model selection through a two-timescale optimization approach, it reduces Quality-of-Service (QoS) violations by up to 1.75× while maintaining high inference accuracy in volatile wireless environments.

What Problem Does This Paper Address?

Offloading heavy machine learning tasks (like real-time video analytics) from mobile devices to edge servers is a standard strategy to save battery and reduce latency. However, two major bottlenecks remain:

  1. Bandwidth Constraints: Transmitting raw sensor data or even intermediate neural features is often too "heavy" for crowded wireless channels.
  2. Channel Volatility: Wireless links suffer from fluctuating Signal-to-Noise Ratios (SNR) and bit errors. Standard compression methods are either too bulky or too fragile—one flipped bit in an entropy-encoded stream can render an entire frame useless for inference.

Existing RAN slicing frameworks often treat the computing task as a static "black box," failing to exploit the flexibility of modern neural architectures that can adapt their complexity and data representation on the fly.

Methods

The authors introduce OMNIS+, which integrates semantic communication directly into the RAN slicing process. Key components include:

  • Dynamic Split Deep Neural Networks (DNNs): Instead of a single model, OMNIS+ uses a multi-branch architecture. A "gate unit" on the mobile device selects a specific branch (varying in feature-channel width) to balance compression ratio against accuracy.
  • Box Quantization: This is a core innovation designed for error resilience. It combines Run-Length Encoding (RLE) to shrink data and Index Embedding (IE)—periodic markers that allow the edge server to resynchronize the bitstream even if some bits are corrupted during transmission.
  • Two-Timescale Optimization:
    • At Task Admission (Mobile Device): MDs use a Causal Contextual Multi-Armed Bandit (MAB) combined with Lyapunov control. This allows the device to learn which DNN branch and edge server to pick based on current channel conditions and queue backlogs, ensuring long-term system stability.
  • Every Time Slot (Edge Server): Servers use Block Coordinate Descent (BCD) to dynamically allocate bandwidth, GPU frequency, and Modulation and Coding Schemes (MCS) among all active tasks.
Fig. 3. System model of OMNIS+. Randomly arriving tasks are queued at each MD. When a task is admitted, the MD selects a dynamic split-DNN branch and a Top-L serving cell and keeps that association until the task fully completes. Concurrent edge-stage tasks share the cell GPU-frequency pool at the co-located BS/ES. Per-cell RAN control allocates bandwidth, GPU frequency, and MCS.

Datasets

The framework was evaluated using the COCO (Common Objects in Context) dataset, specifically for the task of instance segmentation (identifying and outlining individual objects in an image). The authors used a Mask R-CNN architecture with a ResNet-50 backbone as the base for their dynamic split models.

Key Insights and Ideas

  • The Fragility of Efficiency: The authors demonstrate that while entropy-based quantization (like rANS) offers great compression, it is disastrously sensitive to bit errors. In contrast, their "Box" quantization provides a middle ground: better compression than standard INT8 but far more resilience than entropy coding.
  • Semantic-Aware Slicing: Traditional RAN slicing allocates bits. OMNIS+ allocates meaning. By understanding how different compression levels and bit error rates affect final inference accuracy (modeled via Gaussian Processes), the system makes much smarter resource decisions.
  • Joint Optimization is Mandatory: Figure 2 highlights that simply increasing payload size doesn't always improve accuracy, especially at high SNR where models hit a "saturation plateau." The "aha moment" is that the optimal offloading strategy must change based on both the wireless channel and the specific neural branch being used.
Fig. 2. Accuracy vs. payload size under STD quantization at SNR equal to 3 dB (left) and 10 dB (right).

Key Results

  • QoS Reliability: OMNIS+ reduces the probability of missing delay or energy budgets by up to 1.75× compared to greedy optimization baselines.
  • Accuracy Gains: Under poor channel conditions (low SNR), OMNIS+ improves inference accuracy by up to 42.5% over scalar-utility learners by better predicting how different branches will perform.
  • Computational Efficiency: The distributed version of the framework (OMNIS+) achieves nearly the same performance as a centralized controller (C-OMNIS+) but with significantly lower control-plane latency, making it viable for real-time deployment.
Fig. 4. Accuracy vs. payload size for different quantization methods at various SNRs: 3 dB top-left, 5 dB top-right, 7 dB bottom-left, 10 dB bottom-right.

Limitations and Future Work

While OMNIS+ shows robust results for instance segmentation, the paper focuses primarily on computer vision. The authors note that the "Box" quantization parameters (like the scaling factor $\upsilon$) currently require manual tuning for different bottleneck sizes.

Future work could involve:

  • Automating the tuning of quantization parameters through meta-learning.
  • Extending the framework to other modalities like audio or multi-modal sensor fusion.
  • Exploring the impact of more extreme mobility patterns (e.g., high-speed vehicles) on the MAB learning convergence.

Summary updated

Abstract

Edge computing enables resource-constrained devices to execute machine learning (ML) applications via task offloading, with radio access network (RAN) slicing playing a crucial role in provisioning joint communication and computing resources. However, existing RAN slicing frameworks rely on static computing models, limiting their ability to exploit dynamic semantic data representations enabled by modern neural architectures. To address this limitation, we propose OMNIS+, a semantic RAN slicing framework for multi-cell edge computing built upon a new generation of dynamic split neural models. OMNIS+ incorporates dynamic neural compression paired with adaptive quantization for task offloading, providing a versatile decision space across communication payload and compute dimensions. We leveraged a novel ``box'' quantization scheme that markedly enhances resilience to transmission bit errors across diverse compression rates. Considering partial system observability and distinct node objectives, we formulate two interdependent optimization problems: (i)~mobile devices (MDs) maximize inference accuracy under quality-of-service (QoS) and queue-stability constraints by jointly adapting dynamic split deep neural network (DNN) branches and cell association; and (ii)~each edge server (ES) allocates bandwidth, graphics processing unit (GPU) frequency, and modulation and coding scheme (MCS) to maximize the max--min inference accuracy among associated MDs under cell capacity limits. We solve these problems using a multi-agent distributed optimization framework, where MDs act as causal contextual multi-armed bandit (MAB) agents under backlog-aware Lyapunov control, while each ES executes resource allocation via block coordinate descent (BCD). Compared with state-of-the-art baselines, OMNIS+ reduces the QoS violation probability by up to 1.75 while achieving a better accuracy--violation trade-off.
BibTeX
@misc{qin2026omnis,
  author = {Langtian Qin and Avishkar Dnyaneshwar Kawadgave and Ian Andrew Harshbarger and Leïla Nasraoui and Carla Fabiana Chiasserini and Marco Levorato},
  title = {{OMNIS+: Semantic RAN Slicing and Quantization for Dynamic Split Neural Networks}},
  year = {2026},
  eprint = {2611.23977},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.23977},
}
Figure