⚡ electrical engineering

Two-Way Semantic Transmission of Images without Feedback

This paper proposes a two-way semantic communication system that eliminates training feedback overhead by exploiting weight reciprocity and utilizing a conditional generative adversarial network for channel modeling, achieving performance comparable to state-of-the-art one-way systems.

Original authors: Kaiwen Yu, Qi He, Gang Wu

Published 2026-06-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kaiwen Yu, Qi He, Gang Wu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you and a friend are trying to send each other photos of your day, but you're connected by a very noisy, unpredictable walkie-talkie. In the old way of doing this (traditional communication), you'd have to send the photo, wait for your friend to say, "I got a blurry bit here, can you resend?" or "The colors look off," and then you'd adjust. This back-and-forth takes a lot of time and battery power.

In the world of "Semantic Communication" (the new 6G technology), instead of sending raw pixels, you send the meaning of the picture. It's like sending a description of the photo rather than the photo file itself, which is much faster.

However, there's a catch with current smart systems: To teach the computers how to send these "meaning" messages perfectly, they usually need that same annoying back-and-forth feedback loop during training. The receiver has to send error messages back to the transmitter so the transmitter can learn. This creates a heavy traffic jam on the network.

The Paper's Big Idea: The "Mirror" Trick

This paper introduces a new system called TW-SC (Two-Way Semantic Communication) that gets rid of that feedback loop entirely. Here is how they did it, using a few simple analogies:

1. The "Mirror" Effect (Weight Reciprocity)

Imagine you and your friend are standing in front of a perfect mirror. If you raise your right hand, your reflection raises its left. If you learn a dance move, your reflection learns it instantly because you are connected.

The authors realized that in a two-way system (where A talks to B and B talks to A), the connection is often symmetrical (like a mirror). If the "channel" (the walkie-talkie noise) is the same in both directions, the computer brains (neural networks) at both ends can learn the same things.

The Trick: Instead of sending error messages back and forth, Node A can learn how to send better by looking at what Node B would have received, and vice versa. They use a mathematical "mirror" property to train themselves locally without needing to talk to each other during the learning phase.

2. The "Virtual Weatherman" (The Channel Simulator)

To train without feedback, the computers need to know what the "noise" on the line looks like. Usually, they have to wait for the real noise to happen.

The Solution: The authors built a "Virtual Weatherman" using a special AI called a Conditional GAN (Generative Adversarial Network).
How it works: Imagine the AI is a chef trying to guess the recipe of a soup. Usually, the chef has to taste the soup (the real signal) to know if it's good. But here, the chef is given a secret ingredient (a "semantic pilot" or a hint about the message) and uses that to simulate what the soup would taste like in the real world.
This allows the transmitter to practice sending messages through this "Virtual Weatherman" instead of the real, noisy line. It learns to send perfect messages in the simulation, which translates to the real world.

3. The Result: No Traffic Jams

Because the computers can train themselves using the "mirror" trick and the "virtual simulator," they don't need to stop and ask, "Did you get that?" during the training process.

The Outcome: The paper tested this with images (specifically handwritten digits). They found that their new system was almost as good as the best existing systems that do use feedback.
The Benefit: They achieved high-quality image transmission without the heavy "communication overhead" (the extra data needed for feedback). It's like getting a high-definition video call without the lag caused by constant "Can you hear me?" checks.

In Summary
The paper proposes a way for two devices to learn how to send images to each other efficiently by using a mirror-like symmetry and a smart simulator. This allows them to learn the best way to talk without needing to constantly send feedback messages back and forth, saving time and bandwidth while still getting excellent results.

Technical Summary: Two-Way Semantic Transmission of Images without Feedback

Problem Statement
Semantic communications have emerged as a competitive technology for 6G, offering significant improvements in transmission efficiency by recovering the semantic meaning of content rather than just transmitted symbols. While existing research predominantly focuses on one-way (unidirectional) systems, many practical scenarios—such as heterogeneous robot collaboration and cooperative autonomous driving—require bidirectional communication. Current two-way semantic communication systems often rely on Joint Source-Channel Coding (JSCC) trained via deep learning. However, training these systems typically requires back-propagating gradients from the receiver to the transmitter, necessitating information feedback. This feedback introduces substantial communication overhead and latency. Furthermore, offline training strategies suffer performance degradation when the channel conditions during deployment differ from those used during training, while online training incurs additional computational and communication costs.

Methodology
The authors propose a Two-Way Semantic Communication (TW-SC) system designed to eliminate the need for information feedback during the training process. The system operates on two nodes (Node A and Node B) exchanging image data over a reciprocal wireless channel (e.g., Time Division Duplexing).

System Architecture: The system employs a Joint Source-Channel Coding (JSCC) scheme where transmitters consist of semantic and channel encoding neural networks (NNs), and receivers consist of corresponding decoding NNs.
Channel Modeling via SP-CGAN: To simulate the wireless channel without physical transmission during training, the authors utilize a Conditional Generative Adversarial Network (CGAN). Specifically, they introduce a Semantic Pilot-aided CGAN (SP-CGAN). In this model, the received semantic pilot signals are treated as additional conditional information for both the generator and discriminator. This allows the generator to learn the distribution of the wireless channel ( $H$ ) conditioned on the transmitted semantic features.
Weight Reciprocity and Local Training: The core innovation lies in exploiting "weight reciprocity." Assuming reciprocal channels and identical hyperparameters for the NNs at both nodes, the gradients calculated at the receiver are identical to those required at the transmitter. Consequently, the system avoids gradient exchange over the air.
- Stage 1: Transmitters send data to train the channel simulator (SP-CGAN) and the receivers locally.
- Stage 2: Transmitters train their own weights locally. They utilize the trained SP-CGAN to simulate the channel and the locally trained receiver to calculate the loss (Mean Squared Error). The gradients are back-propagated internally within the node, eliminating the need for feedback from the remote node.

Key Contributions

Feedback-Free Training: The paper presents a TW-SC framework where transceiver NNs are trained locally without requiring gradient feedback between nodes, significantly reducing communication overhead and delay.
SP-CGAN Channel Simulator: The introduction of a channel simulator that incorporates semantic pilots as conditional information improves the accuracy of channel distribution modeling compared to standard GANs.
Local Optimization: By leveraging weight reciprocity, the proposed method allows transmitters to optimize their parameters using internal loss calculations derived from the simulated channel and local receiver, applicable to both online and offline scenarios.

Numerical Results
The proposed TW-SC system was evaluated using the MNIST handwritten digit dataset under Additive White Gaussian Noise (AWGN) and Rayleigh fading channels. It was compared against two benchmarks:

Optimal One-way JSCC: Trained on real channels (serving as the performance upper bound).
GAN-SC: A one-way system using a standard GAN for channel simulation without semantic pilots.

Performance: In AWGN channels, the TW-SC system (TW-SC-AWGN) achieved performance close to the optimal JSCC, particularly at low Signal-to-Noise Ratios (SNR). In Rayleigh fading channels, the TW-SC trained under Rayleigh conditions outperformed models trained under AWGN at lower SNRs, demonstrating robustness.
Metrics: The system was evaluated using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM). The TW-SC scheme consistently outperformed the GAN-SC benchmark, which suffered from poor performance due to the lack of semantic pilot conditioning.
Convergence: Training curves showed that the TW-SC links (A to B and B to A) converged within 20 epochs, with performance trajectories nearly overlapping with the optimal JSCC algorithm.

Significance and Claims
The paper claims that the proposed TW-SC system effectively bridges the gap between high-performance one-way semantic communication and the requirements of bidirectional tasks. By removing the feedback overhead typically associated with JSCC training, the system offers a more practical and efficient solution for 6G applications. The authors state that their system performs "closing to" (close to) state-of-the-art one-way systems while requiring no feedback between transceivers during training. The work suggests that this framework can be extended to other data types beyond images, provided the channel reciprocity and weight symmetry assumptions hold.

1. The "Mirror" Effect (Weight Reciprocity)

2. The "Virtual Weatherman" (The Channel Simulator)

3. The Result: No Traffic Jams

More like this