Abstract:State-of-the-art performance for many edge applications is achieved by deep neural networks (DNNs). Often, these DNNs are location- and time-sensitive, and must be delivered over a wireless channel rapidly and efficiently. In this paper, we introduce AirNet, a family of novel training and transmission methods that allow DNNs to be efficiently delivered over wireless channels under stringent transmit power and latency constraints. This corresponds to a new class of joint source-channel coding problems, aimed at delivering DNNs with the goal of maximizing their accuracy at the receiver, rather than recovering them with high fidelity. In AirNet, we propose the direct mapping of the DNN parameters to transmitted channel symbols, while the network is trained to meet the channel constraints, and exhibit robustness against channel noise. AirNet achieves higher accuracy compared to separation-based alternatives. We further improve the performance of AirNet by pruning the network below the available bandwidth, and expanding it for improved robustness. We also benefit from unequal error protection by selectively expanding important layers of the network. Finally, we develop an approach, which simultaneously trains a spectrum of DNNs, each targeting a different channel condition, resolving the impractical memory requirements of training distinct networks for different channel conditions.

What problem does this paper attempt to address?

### Problems the Paper Attempts to Solve This paper aims to address the issue of reliably transmitting machine learning models over wireless links under strict bandwidth and transmission power constraints, particularly for efficient inference in edge computing scenarios. Specifically, the authors propose a novel method called AirNet, which is a Joint Source-Channel Coding (JSCC) approach that enables effective transmission of deep neural network (DNN) parameters under low Signal-to-Noise Ratio (SNR) and limited bandwidth conditions. #### Major Contributions Include: 1. **Proposing a New Communication Problem**: Efficiently transmitting machine learning models over wireless links under strict bandwidth and transmission power constraints to achieve reliable inference at the receiver end. 2. **Introducing the AirNet Architecture**: A new JSCC method that achieves high-precision inference under low SNR and bandwidth-limited conditions, and is more robust compared to separate methods. 3. **Network Pruning and Knowledge Distillation**: Utilizing network pruning to meet bandwidth constraints and using knowledge distillation to improve inference accuracy at the receiver end. Adding noise during training further enhances robustness. 4. **Unequal Error Protection (UEP)**: Selectively protecting important layers through bandwidth expansion techniques, significantly improving the final reception accuracy. 5. **Multi-Condition Adaptation**: Proposing an ensemble learning method that trains multiple networks for different channel conditions simultaneously, addressing the issue of needing to store different network parameters for different channel conditions. ### Research Background With the development of deep learning, complex deep neural networks have shown excellent performance in various applications, but their large model sizes pose challenges for wireless transmission. Traditional separate methods (compress first, then encode and transmit) perform poorly in low SNR environments, whereas AirNet maps DNN parameters directly to channel symbols and considers the impact of channel noise during training, achieving better performance. Additionally, techniques such as pruning, knowledge distillation, and bandwidth expansion further enhance the system's robustness and accuracy.

AirNet: Neural Network Transmission over the Air

M-Net: A Lightweight Network Based on Multilayer Perceptron for Massive MIMO CSI Feedback

Implementing Graph Neural Networks over Wireless Networks Via Over-the-Air Computing: A Joint Communication and Computation Framework

AirNN: Neural Networks with Over-the-Air Convolution via Reconfigurable Intelligent Surfaces

Federated AirNet: Hybrid Digital-Analog Neural Network Transmission for Federated Learning

Graph Neural Networks over the Air for Decentralized Tasks in Wireless Networks

Over-the-Air Computation Empowered Vertically Split Inference

Deep Learning-Based Communication Over the Air

NetNN: Neural Intrusion Detection System in Programmable Networks

Communication-Efficient Separable Neural Network for Distributed Inference on Edge Devices

In-Network Learning Via Over-the-Air Computation in Internet-of-Things

Utilizing Multiple-Access Communication to Realize FNN Computation in Wireless Networks

AirNet: Fused Analytical and Iterative Reconstruction with Deep Neural Network Regularization for Sparse-Data CT

AirNet: Fused Analytical and Iterative Image Reconstruction Method with Deep Learning Regularization for High-Quality Sparse-Data On-Board CBCT

Edge-Host Partitioning of Deep Neural Networks with Feature Space Encoding for Resource-Constrained Internet-of-Things Platforms

A Learning-Based End-to-End Wireless Communication System Utilizing a Deep Neural Network Channel Module

NNCTC: Physical Layer Cross-Technology Communication via Neural Networks

Wireless Neural Network: Enabling Neural Computing Over Wireless Sensor Network Based On Superposition Transmissions

NeuralEQ: Neural-Network-Based Equalizer for High-Speed Wireline Communication

Deep Neural Networks with Data Rate Model: Learning Power Allocation Efficiently

Resource-Efficient Distributed Deep Neural Networks Empowered by Intelligent Software-Defined Networking.