Distributed Deep Joint Source-Channel Coding over a Multiple Access Channel

Selim F. Yilmaz, Can Karamanli, Deniz Gündüz

TL;DR

This paper introduces DeepJSCC-NOMA, a deep learning framework that allows multiple devices to transmit images simultaneously over a shared wireless channel. By combining Non-Orthogonal Multiple Access (NOMA) with joint source-channel coding, the researchers achieved significantly better image reconstruction quality than traditional "take-turns" (TDMA) methods, especially when bandwidth is scarce.

What Problem Does This Paper Address?

In traditional wireless communication, we separate source coding (compressing the data) from channel coding (adding redundancy to protect against noise). While Shannon’s theory says this separation is optimal in a perfect world with infinite data, it fails in the real world where we have finite-length data (like a single image) and low signal-to-noise ratios (SNR).

The problem gets even harder in a Multiple Access Channel (MAC), where multiple devices want to talk to one receiver. Usually, we use "orthogonal" methods like TDMA, where users take turns. However, information theory suggests that "non-orthogonal" methods—where everyone talks at once and the receiver untangles the mess—can be much more efficient. The challenge is: how do you design a practical system that can "untangle" complex data like images transmitted simultaneously?

Methods

The authors propose a novel architecture called DeepJSCC-NOMA. Key technical components include:

  • Deep Joint Source-Channel Coding (DeepJSCC): Instead of separate compression and error correction, a single Deep Neural Network (DNN) maps image pixels directly to complex-valued channel symbols.
  • Siamese Network Architecture: To keep the system scalable, both transmitters use the exact same encoder parameters. This prevents the "parameter explosion" that usually happens as you add more users.
  • Device Embeddings: Since both encoders are identical, the receiver needs a way to tell the signals apart. The authors introduce "device embeddings"—unique, trainable weights concatenated to the input image. This acts like a digital "fingerprint" that helps the decoder separate the superposed signals.
  • Curriculum Learning (CL): Training a network to decode two interfering signals at once is hard. The authors use a "progressive" training strategy: first, they train the network to transmit images without interference (an easier task), then they fine-tune it on the full NOMA task.
  • Power Normalization: A dedicated layer ensures each transmitter stays within its power budget.

Datasets

The researchers evaluated their method using the CIFAR-10 dataset, a standard benchmark in computer vision consisting of 60,000 small (32x32) RGB images across 10 classes. They used 45,000 images for training, 5,000 for validation, and 10,000 for testing.

Key Insights and Ideas

  • Superposition is Superior: The "aha moment" is that letting signals interfere and then using a smart DNN to separate them (NOMA) yields better results than splitting the available time/bandwidth between users (TDMA).
  • The "Identity" Problem: In NOMA, the receiver doesn't just need to reconstruct the data; it needs to know whose data it is. The use of simple 1D device embeddings is a remarkably low-overhead way to solve this, adding only about 1% more parameters to the model.
  • Graceful Degradation: Unlike traditional digital systems that "break" completely (the cliff effect) if the noise gets too high, this deep learning approach exhibits "graceful degradation"—the image quality drops slowly as the channel worsens, but it remains decodable.
  • Fairness: Despite the complexity of decoding two signals at once, the system is "fair," meaning both users achieve nearly identical image quality (PSNR).

Key Results

The authors compared DeepJSCC-NOMA against a TDMA-based DeepJSCC baseline across various Signal-to-Noise Ratios (SNR) and bandwidth ratios ($\rho$):

  • Higher Quality: For a bandwidth ratio of $\rho = 1/3$, the NOMA method with Curriculum Learning achieved a 0.91 dB improvement in Peak Signal-to-Noise Ratio (PSNR) over the TDMA baseline.
  • Low Bandwidth Efficiency: At a tighter bandwidth ratio of $\rho = 1/6$, the improvement was 0.25 dB.
  • Efficiency: The Siamese architecture allows the system to scale to multiple users with only a marginal increase in computational and memory complexity compared to a single-user system.

Limitations and Future Work

While the results are promising, there are a few areas for further exploration:

  • Scaling Up: The experiments were limited to two users and small 32x32 images. Testing on higher-resolution datasets (like ImageNet) and with more simultaneous users would be the next logical step.
  • Channel Assumptions: The study assumes an Additive White Gaussian Noise (AWGN) channel. Real-world wireless channels involve fading and multi-path interference, which would require more robust model training.
  • Synchronization: The paper assumes the transmitters are perfectly synchronized. In practice, managing timing offsets between multiple devices in a NOMA system is a significant engineering challenge.

Summary updated

Abstract

We consider distributed image transmission over a noisy multiple access channel (MAC) using deep joint source-channel coding (DeepJSCC). It is known that Shannon's separation theorem holds when transmitting independent sources over a MAC in the asymptotic infinite block length regime. However, we are interested in the practical finite block length regime, in which case separate source and channel coding is known to be suboptimal. We introduce a novel joint image compression and transmission scheme, where the devices send their compressed image representations in a non-orthogonal manner. While non-orthogonal multiple access (NOMA) is known to achieve the capacity region, to the best of our knowledge, non-orthogonal joint source-channel coding (JSCC) scheme for practical systems has not been studied before. Through extensive experiments, we show significant improvements in terms of the quality of the reconstructed images compared to orthogonal transmission employing current DeepJSCC approaches particularly for low bandwidth ratios. We publicly share source code to facilitate further research and reproducibility.
BibTeX
@inproceedings{yilmaz2023distributed,
  author = {Selim F. Yilmaz and Can Karamanli and Deniz Gündüz},
  title = {{Distributed Deep Joint Source-Channel Coding over a Multiple Access Channel}},
  booktitle = {ICC},
  year = {2023},
  doi = {10.1109/icc45041.2023.10279499},
}

Loading related content...

Figure