Semantic Communications With Discrete-Time Analog Transmission: A PAPR Perspective

Yulin Shao, Deniz Gündüz

This paper, authored by Yulin Shao and Deniz Gündüz from Imperial College London, tackles a major practical hurdle in the deployment of next-generation semantic communication systems: the "Power Efficiency Problem."

TL;DR

Deep learning-based semantic communication (DeepJSCC) offers superior image transmission but suffers from a high Peak-to-Average Power Ratio (PAPR), which drains mobile batteries and reduces coverage. This research demonstrates that "trainable clipping" can slash this power efficiency bottleneck by over 15 dB with minimal impact on image quality, proving that semantic gains don't depend on high-power signal peaks.

What Problem Does This Paper Address?

Traditional digital communications use discrete "constellations" (like QAM) to send data. Modern semantic communication, specifically Deep Joint Source-Channel Coding (DeepJSCC), throws out the rulebook by using Discrete-Time Analog Transmission (DTAT). Instead of bits, it sends continuous-valued features directly from a neural network.

While DeepJSCC is incredibly robust against noise, it has a "spiky" signal profile. When combined with OFDM (the modulation used in 4G/5G), it creates a very high Peak-to-Average Power Ratio (PAPR). High PAPR is a nightmare for hardware: it forces power amplifiers to operate inefficiently to avoid distorting the signal peaks. For a mobile user, this means shorter battery life and poor uplink performance.

The authors set out to answer a critical question: Do the performance gains of DeepJSCC come from the freedom of these high-power peaks, or can we suppress them without losing the "semantic" edge?

Methods

The authors developed a passband transceiver model to accurately measure PAPR (most previous studies only looked at baseband, which underestimates the problem). They evaluated three main PAPR reduction strategies:

  1. Linearly Precoded OFDMA: Using techniques like LFDMA and IFDMA (similar to the "Single Carrier" approach used in 4G LTE uplink) to spread the signal energy.
  2. PAPR Loss Function: Adding a penalty term directly into the neural network's loss function during training, essentially telling the AI: "Minimize reconstruction error, but also keep the signal peaks low."
  3. Trainable Clipping: This is the most novel approach. The authors implemented a differentiable clipping operation within the end-to-end training loop. By simulating signal clipping during training, the DeepJSCC "learns" to be robust to the resulting interference.

Datasets

The study utilized three standard computer vision datasets:

  • CIFAR10: Used for initial training and characterizing the trade-off curves (ROC) between power and image quality.
  • ImageNet: Used for large-scale training (1.2 million images) to ensure the findings generalize to complex, high-resolution data.
  • Kodak: A standard set of 24 high-quality images used for final testing and benchmarking against traditional codecs.

Key Insights and Ideas

  • The "Freedom" Myth Debunked: The researchers confirmed that the "magic" of DeepJSCC isn't just a byproduct of high-PAPR signals. You can aggressively constrain the signal's dynamic range and still maintain a massive performance lead over traditional digital methods (like BPG compression + LDPC coding).
  • Deep Learning as a "Cleaner": Traditional receivers struggle with the Inter-Carrier Interference (ICI) caused by clipping. However, because DeepJSCC is trained end-to-end, the decoder learns to "denoise" the specific distortions caused by its own transmitter's clipping.
  • Clipping is King: While linear precoding (IFDMA) helps, it isn't enough on its own for analog-like signals. Trainable clipping emerged as the most effective tool for balancing power efficiency and image reconstruction.

Key Results

  • Massive PAPR Reduction: Without intervention, DeepJSCC had a staggering PAPR of ~21 dB. Using trainable clipping, the authors reduced this to 6 dB—a level acceptable for practical cellular hardware.
  • Minimal Quality Loss: Reducing the PAPR by 15 dB resulted in a Peak Signal-to-Noise Ratio (PSNR) drop of only about 2-3 dB.
  • Still Beating the Benchmarks: Even with the 6 dB PAPR constraint, DeepJSCC achieved a PSNR of ~27.4 dB at a 10 dB SNR, significantly outperforming the best traditional digital benchmark (BPG + 16QAM + LDPC), which topped out at ~22.3 dB.

Limitations and Future Work

  • Channel Specifics: The paper focuses primarily on the AWGN (Additive White Gaussian Noise) channel. While the authors suggest the results will hold for fading channels, more complex multi-path environments remain to be fully explored.
  • Hardware Realities: The study is simulation-based. The next step involves testing these "low-PAPR" semantic signals on real-world Software Defined Radios (SDRs) to measure actual power consumption and heat dissipation in the amplifiers.
  • Computational Overhead: Incorporating clipping and PAPR penalties into the training process increases the complexity of the training phase, though it does not affect the speed of the transmitter during actual use.

Summary updated

Abstract

Recent progress in deep learning (DL)-based joint source-channel coding (DeepJSCC) has led to a new paradigm of semantic communications. Two salient features of DeepJSCC-based semantic communications are the exploitation of semantic-aware features directly from the source signal, and the discrete-time analog transmission (DTAT) of these features. Compared with traditional digital communications, semantic communications with DeepJSCC provide superior reconstruction performance at the receiver and graceful degradation with diminishing channel quality, but also exhibit a large peak-to-average power ratio (PAPR) in the transmitted signal. An open question has been whether the gains of DeepJSCC come at the expense of high-PAPR continuous-amplitude signal, which can limit its adoption in practice. In this letter, we first show that conventional DeepJSCC does suffer from high PAPR. Then, we explore three PAPR reduction techniques and confirm that the superior image reconstruction performance of DeepJSCC can be retained while the PAPR is suppressed to an acceptable level. This is an important step towards the implementation of DeepJSCC in practical semantic communication systems.
BibTeX
@article{shao2023semantic,
  author = {Yulin Shao and Deniz Gündüz},
  title = {{Semantic Communications With Discrete-Time Analog Transmission: A PAPR Perspective}},
  journal = {IEEE Wireless Communications Letters},
  year = {2023},
  doi = {10.1109/lwc.2022.3232946},
}

Loading related content...

Figure