Learning Multi-Rate Task-Oriented Communications Over Symmetric Discrete Memoryless Channels

Anbang Zhang,Shuaishuai Guo

2024-08-24

Abstract:This letter introduces a multi-rate task-oriented communication (MR-ToC) framework. This framework dynamically adapts to variations in affordable data rate within the communication pipeline. It conceptualizes communication pipelines as symmetric, discrete, memoryless channels. We employ a progressive learning strategy to train the system, comprising a nested codebook for encoding and task inference. This configuration allows for the adjustment of multiple rate levels in response to evolving channel conditions. The results from our experiments show that this system not only supports edge inference across various coding levels but also excels in adapting to variable communication environments.

Signal Processing,Image and Video Processing

What problem does this paper attempt to address?

### Problems the paper attempts to solve

This paper aims to solve the problem that in a dynamic communication environment, traditional task - oriented communication (ToC) systems are unable to adapt to changing data transmission rates. Specifically:

1. **Limitations of traditional communication systems**:
- Traditional communication systems rely on accurate bit recovery and require good channel conditions and sufficient communication resources.
- These systems are designed for general - purpose data transmission and do not specifically consider the needs or requirements of the current task.
- In cases where time - sensitivity and data - relevance are crucial, traditional static models often fail to meet the requirements.

2. **The need for multi - rate task - oriented communication**:
- As the communication environment changes, such as the modulation order change in the wireless channel, the existing ToC framework needs to adjust the encoder, task - inference module, and codebook.
- This inflexibility leads to problems with the integrity and reliability of remote - inference tasks.

3. **Proposed solutions**:
- The authors introduce an adaptive multi - rate task - oriented communication (MR - ToC) framework.
- By using nested codebooks and a progressive learning strategy, this framework can be flexibly adjusted at different transmission rates and only requires one encoder, one task - inference module, and one nested codebook.

4. **Specific objectives**:
- Develop a general - purpose adaptive system that can operate efficiently at all potential transmission rates.
- Improve the performance of edge - inference tasks, especially in a dynamic communication environment.

### Formula representation

The formulas involved in the paper are as follows:

- **Nested codebook mapping**:
\[
z_c(x)=\text{argmin}_{e_j}\|z_e(x) - e_j\|^2,\quad\forall e_j\in Q
\]
where \(z_c(x)\) and \(z_e(x)\) represent the discretized vector and the continuously represented sub - vector respectively.

- **Probability transition matrix of the SDMC model**:
\[
P=\begin{bmatrix}
1-\epsilon&\frac{\epsilon}{r - 1}&\cdots&\frac{\epsilon}{r - 1}\\
\frac{\epsilon}{r - 1}&1-\epsilon&\cdots&\frac{\epsilon}{r - 1}\\
\vdots&\vdots&\ddots&\vdots\\
\frac{\epsilon}{r - 1}&\frac{\epsilon}{r - 1}&\cdots&1-\epsilon
\end{bmatrix}_{r\times r}
\]
where \(\epsilon\) represents the transmission error probability and \(r\) is the total number of codewords used.

- **Loss function**:
\[
L^{(l)}_{\text{MR - ToC}}(y_t;x_t)=\sum_{j = 1}^l L(y_t;\hat{y}_t^{(j)})+\| \text{sg}[z_e,t(x)]-e_j\|^2_2+\gamma_j\|z_e,t(x)-\text{sg}[e_j]\|^2_2+\eta_l\sum_{k = 1}^{2^{l-1}}\|e_k^{(l)}-e_k^{(l-1)}\|^2_2
\]

Through these formulas, the paper shows how to achieve the effectiveness of multi - rate task - oriented communication through progressive learning and nested codebooks.