Rateless Joint Source-Channel Coding, and a Blueprint for 6G Semantic Communications System Design

TL;DR

The code is rateless in that it is designed and optimized for a continuum of coding rates such that it achieves a desired distortion for any rate in that continuum, and a new family of autoencoder rateless JSCC codes are introduced.

Abstract

This paper introduces rateless joint source-channel coding (rateless JSCC). The code is rateless in that it is designed and optimized for a continuum of coding rates such that it achieves a desired distortion for any rate in that continuum. We further introduce rate-adaptive and stable communication link operation to accommodate rateless JSCCs. The link operation resembles a ``bit pipe'' that is identified by its rate in bits per frame, and, by the rate of bits that are flipped in each frame. Thus, the link operation is rate-adaptive such that it punctures the rateless JSCC codeword to adapt its length (and coding rate) to the underlying channel capacity, and is stable in maintaining the bit flipping ratio across time frames. Next, a new family of autoencoder rateless JSCC codes are introduced. The code family is dubbed RLACS code (read as relax code, standing for ratelss and lossy autoencoder channel and source code). The code is tested for reconstruction loss of image signals and demonstrates powerful performance that is resilient to variation of channel quality. RLACS code is readily applicable to the case of semantic distortion suited to variety of semantic and effectiveness communications use cases. In the second part of the paper, we dive into the practical concerns around semantic communication and provide a blueprint for semantic networking system design relying on updating the existing network systems with some essential modifications. We further outline a comprehensive list of open research problems and development challenges towards a practical 6G communications system design that enables semantic networking.

Figures (10)

• Figure 1: Rateless JSCC and rate-adaptive link for semantic communication. At the bottom, $\mathbf{x}$ and $\mathbf{y}$ are shown as binary sequences, as an example. In general, the symbols selected for communication are chosen from a constellation based on the estimated channel knowledge at the transmitter. The bits in red represent the flipped bits due to communication error, while the shaded blocks represent the null bits at the decoder caused by puncturing in the network.
• Figure 2: Training iteration $i$ for RLACS code, where the trained encoder and decoder modules from previous iterations are frozen (blue color). Dashed arrow lines point to optimization of $f_i$ and $g_i$ based on the loss calculation.
• Figure 3: Architecture of the used for $f_{\theta_i}$ and $g_{\phi_i}$ are shown on the right and the left sides, respectively. The dimensions of the encoder and decoder layers adapts to $C_i - C_{i-1}$ and $C_i$, to ensure a sufficient capacity in the network architecture. Additionally, in the decoder structure, layer dimensions first expand through the middle of the network and then contract, using a factor $z = \min (1, 256/C_i)$. To simplify the illustration, a "Big Block" is introduced in the bottom right corner, which takes a variable $x$ to choose between up-sampling and down-sampling (stride 2), if either is used. In the decoder, skip connections retain intermediate outputs from earlier layers and integrate them into later layers. The retained outputs are resized to match the current layer's resolution and combined with its features. An attention module then selects the most important information from the combined features.
• Figure 4: Average bit error rate experienced over the testing channels $\text{SVBSC}(q_o=0.05,\epsilon, 1280, 1152)$ (top legend) against the underlying wireless channel with $\bar{W}=128$. The uncoded QAM curves cover $M = {2, 4, 8, 16, 32, 64, 128, 256, 512, 1024}$ modulation orders (ordered from left to right). In the case of perfect , can be perfectly kept below $q_o$. In case of imperfect , $\epsilon$ determines the average . A tight $\epsilon$ of $0.01$ for instance, results in average almost one order of magnitude below $q_o$. Average is not in itself an important measure---more critical metric is the average reconstruction loss over the randomness of , when is imperfect.
• Figure 5: Average spectral efficincy of the curves in Fig. \ref{['fig:bervssnr']}. With $\epsilon = 0.05$ the average spectral efficiency, and the average closely follow the case of perfect .

Theorems & Definitions (10)

• Remark 1
• Remark 2
• Remark 3
• Definition 1
• Remark 4
• Remark 5
• Definition 2
• Definition 3
• Remark 6
• Remark 7