Discrete-Space Generative AI Pipeline for Semantic Transmission of Signals

1. Problem Statement

The paper addresses limitations in current Semantic Communication (SC) systems, which aim to transmit the meaning of a message rather than raw bits. Key challenges identified include:

• Domain Specificity: Existing SC use cases are often limited to specific domains (text, vision, speech) and lack generalization to diverse physical signals like radio frequency (RF) and general audio.
• Computational Complexity: Generative models used for SC are often computationally expensive and power-hungry, making them unsuitable for resource-constrained IoT devices.
• Channel Robustness: Classical SC systems struggle when the physical channel introduces severe impairments (e.g., packet loss, truncation, or erasures), leading to a loss of semantic integrity.
• Continuous vs. Discrete: Most generative models operate in continuous latent spaces, which can be inefficient for transmission. The paper proposes exploring discrete latent spaces to reduce complexity and improve robustness.

2. Methodology: The Discernment Framework

The authors propose Discernment, a semantic communication system that operates in a discrete latent space. The pipeline consists of four main stages:

A. Discretization (VQ-VAE)

• Input: High-dimensional physical signals (Baseband RF IQ samples or Audio Spectrograms).
• Process: A Vector Quantized Variational Autoencoder (VQ-VAE) maps continuous signals into a sequence of discrete tokens (codebook indices).
• Mechanism: The encoder quantizes the input into the nearest codeword from a learned codebook. The output is a sequence of integers ($z_Q$) representing the signal's latent structure.
• Benefit: This simplifies the generative modeling task, reduces memory requirements, and transforms the problem into sequence generation.

B. Generative Modeling (GenAI)

The system employs two distinct Generative AI architectures to model the probability distribution of the discrete tokens, selected based on the expected channel impairment pattern:

1. Decoder-only Transformers (DoT):
   • Architecture: Lightweight autoregressive models (e.g., NanoGPT, MONAI).
   • Use Case: Optimized for Truncation Channels (where the end of the sequence is missing).
   • Mechanism: Predicts tokens sequentially ($z_{t+1}$ given $z_{1...t}$).
2. Score Entropy Discrete Diffusion (SEDD):
   • Architecture: A non-autoregressive diffusion model adapted for discrete data.
   • Use Case: Optimized for Puncturing/Erasure Channels (where tokens are randomly missing throughout the sequence).
   • Mechanism: Uses a "Mask-and-Replace" strategy. It learns to reconstruct masked tokens in parallel by reversing a noise process where tokens are replaced by a special mask token.

C. Semantic Channel Model

• The technical channel is modeled as an Erasure Channel.
• Transmission: Only a subset of the discrete tokens is transmitted.
• Recovery: The receiver uses the appropriate GenAI model (DoT or SEDD) to infer the missing tokens (erasures) based on the received context and the learned prior distribution.
• Reconstruction: The completed token sequence is fed into the VQ-VAE decoder to reconstruct the original physical signal (RF waveform or audio spectrogram).

D. Evaluation Metric

Instead of relying solely on signal reconstruction fidelity (PSNR/MSE), the system is evaluated based on Task Alignment. A pre-trained classifier (CNN) evaluates the reconstructed signals to determine if the "meaning" (e.g., modulation type or spoken digit) is preserved despite data loss.

3. Key Contributions

1. Discrete-Space Semantic Communication: The first comparative analysis of transformer-based and diffusion-based generative models operating on discrete latent representations for both RF and audio signals.
2. Adaptive Architecture: A novel framework that dynamically switches between autoregressive (DoT) and diffusion-based (SEDD) generation strategies depending on the specific pattern of channel erasures (truncation vs. puncturing).
3. SEDD for Non-Language Signals: Adapting the Score Entropy Discrete Diffusion (SEDD) model, previously used for language, to generate diverse physical signals (RF and audio) with high efficiency.
4. Lightweight Design: Demonstrating that semantic communication can be achieved with lightweight models suitable for edge/IoT deployment, reducing computational complexity compared to continuous-space GANs or hybrid architectures.

4. Experimental Results

The system was tested on two datasets: TorchSig (6 RF modulation classes) and AudioMNIST (spoken digits 0-9).

• Robustness to Erasures:
  • DoT (Truncation): Maintained high classification accuracy even when up to 87.5% of the sequence was truncated (only 12.5% of tokens received).
  • SEDD (Puncturing): Showed remarkable robustness to random packet loss. Even with only 3% of tokens received (97% erasure), classification accuracy dropped by only 1.07% (remaining at ~98.93%).
• Performance Comparison:
  • SEDD vs. DoT: SEDD demonstrated superior robustness to random erasure patterns compared to DoT, which relies on sequential context.
  • Codebook Size: Interestingly, larger codebook sizes ($N$) allowed for earlier recovery of task-aligned meaning, contrary to the intuition that larger vocabularies increase uncertainty.
• Efficiency: The models achieved near-perfect classification accuracy (100% on clean data) with significantly lower parameter counts compared to state-of-the-art continuous generative models (e.g., Jukebox, HuBERT).

5. Significance and Future Work

• Paradigm Shift: Discernment validates the "semantic-channel" paradigm where the channel is treated as an erasure mechanism and GenAI acts as the error correction code, preserving semantic integrity even under severe bandwidth constraints.
• IoT Applicability: The low model complexity and ability to operate on discrete tokens make this framework highly suitable for resource-constrained IoT devices that need to transmit sensor data to more powerful edge/cloud nodes.
• Future Directions:
  • Investigating the scaling of robustness against latent compression rates.
  • Incorporating contrastive learning before the VQ-VAE to cluster semantically similar content in the latent space, potentially improving the completion of partially received descriptions.
  • Exploring the trade-off between computational complexity and task complexity.

In conclusion, Discernment demonstrates that by leveraging discrete latent spaces and adaptive generative AI, semantic communication systems can achieve graceful degradation and high fidelity in meaning transmission, even when the physical channel is severely impaired.