Anchor-Aided Multi-User Semantic Communication with Adaptive Decoders
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.
Tom: We're looking at the paper: "Anchor-Aided Multi-User Semantic Communication with Adaptive Decoders".
Jane: Here's the gist of the paper. Semantic communication (SemCom) can actively embed the semantic meaning of the data into the transmission process, while eliminating statistical redundancy to preserve bandwidth resources for other users Most existing works only consider one transmitter and one receiver, which limits their ability to address the diversity in users’ models and capabilities Therefore, in this paper, we propose a multi-user semantic communication system where each user is equipped with a distinct DL-based joint source-channel decoder architecture, reflecting the diversity in computing capacity The challenging issue with the proposed system is the catastrophic forgetting property of neural networks, where the DL-based encoder fails to encode the data for the previous user when being trained with a new user.
Tom: System Architecture and Heterogeneity The proposed system involves a downlink transmission from a base station (BS) to multiple users, each equipped with a distinct DL-based decoder architecture tailored to its computational capability This cross-architecture poses a critical challenge for training DL models in the network: the shared BS encoder received different updated directions from clients in the network, which leads to the catastrophic forgetting.
Jane: The system considers users architecturally heterogeneous, reflecting the asynchronous and resource-constrained nature of practical devices in a real-world network.
Tom: Anchor-Aided Training Strategy To address the gradient conflict problem, a novel two-stage training framework is proposed that decouples encoder optimization from decoder adaptation In the first stage, the encoder is trained together with a decoder that shares a symmetric architecture, which promotes stable learning and allows the encoder to fully capture semantic representations.
Title and authors: Tom: So, let's talk about the title and the people behind this work—"Anchor-Aided Multi-User Semantic Communication with Adaptive Decoders." What does that actually mean in plain English?
Jane: It means they are using a method called anchor-aided training to make sure each user gets a decoder that fits their specific needs, which is key for handling multiple users.
Lu: The authors are X. Nguyen, Phuong-Nam Tran, Trung Thanh Pham, Avi Deb Raha, Eui-Nam Huh, Zhu Han, and Choong Seon Hong. They bring expertise in the deep learning architectures used here and the overall communication system design.
Meng: I'm curious about the "anchor-aided" part. Does that mean they have some kind of fixed reference point to guide how those decoders adapt? Because without a fixed point, training all those different models would probably just get messy.
Lalam: The anchor is basically something that helps the system maintain stability while allowing each user's decoder to learn its own optimal way to interpret the data. It’s about guiding the learning process in a structured way.
Tom: It sounds like they’re trying to solve a real problem where you have this massive diversity in hardware capabilities, and they’re proposing an adaptive solution instead of a one-size-fits-all approach.
The paper's summary: Jane: Okay, so the paper summarizes this whole system by saying that semantic communication can actively embed the data's meaning into the transmission while cutting out statistical redundancy to save bandwidth for everyone else.
Tom: And they argue that most existing systems are limited because they only look at a single transmitter and receiver, which restricts how well we can manage different user models.
Lu: They propose this multi-user system with distinct DL based joint source-channel decoders for each user, which directly reflects the real world where devices have very different computing capacities.
Meng: The summary says they are focusing on the D-JSCC direction of semantic communication and addressing a key challenge through an adaptive training strategy rather than just any random model change.
Lalam: They introduce a two-stage training framework to handle the issue of conflicting updates that happens when you try to train these diverse decoders simultaneously against one shared encoder.
The paper's improvements: Tom: What’s the actual improvement they’re pushing? It sounds like they tackle the problem of catastrophic forgetting, which is a big deal in continual learning scenarios.
Jane: That’s right. They show that when you train the encoder with a new user, it doesn't forget what it learned for previous users, which is a major hurdle for these types of systems.
Lu: They solve this by using an anchor-aided two stage training framework that decouples the encoder optimization from the decoder adaptation process. This keeps the representations stable while letting them adapt to new users efficiently.
Meng: So, they train a shared architecture with a symmetric decoder first, and then freeze the encoder parameters before training decoders for different users separately? That sounds like it simplifies things for deployment.
Lalam: Precisely. The paper shows that after that initial self reflective training stage, the optimization problems for each user decoder become independent, meaning there's no gradient interference happening across users anymore.
Conclusion: Tom: So, to wrap up this whole discussion on "Anchor-Aided Multi-User Semantic Communication with Adaptive Decoders," they show that their framework outperforms conventional training schemes across a wide range of channel qualities.
Jane: They achieve the highest performance across all types of decoder users and a wide range of channel qualities compared to iterative and simultaneous training cases.
Lu: The reason for this high performance is that the symmetric training successfully guides the base station encoder to extract data meaning in a way that compresses it meaningfully, which is pretty smart representation learning.
Meng: For production, what I’m really interested in is how robust this setup actually is when we run it on real traffic—does it handle sudden changes in user density or model complexity well?
Lalam: The authors show that this framework achieves higher performance across almost all levels of channel noise in both channel models, which suggests a solid foundation for real-world deployment.
Tom: So, the big picture is that they’ve successfully navigated the tricky gradient conflict problem and created a way to handle diverse user needs in semantic communication. We'll keep an eye on how this adaptive training strategy scales.
Kyung Hee University Department of Electrical and Computer Engineering University of Houston Department of Computer Science and Engineering
cs.IT, cs.ET, math.IT
Submitted: 2026-04-14
Updated: 2026-05-08
Comments: 15 pages, 7 figures
Journal ref: IEEE Transactions on Network Science and Engineering 2026
DOI: 10.1109/TNSE.2026.3750303
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 81/100
The gist: Semantic communication (SemCom) can actively embed the semantic meaning of the data into the transmission process…
Key concepts
- Semantic Communication (SemCom)
- A communication technique that embeds the actual meaning of the data directly into the transmission process. It achieves this by removing statistical redundancy, allowing more bandwidth to be used for other users while preserving essential information.
- Architectural Heterogeneity
- The system accommodates users with different deep learning decoder architectures. This reflects real-world constraints where devices have varying computational capacities, allowing the network to serve diverse client models simultaneously.
- Catastrophic Forgetting
- A problem where a neural network trained on one task (or user) loses its ability to perform well on previously learned tasks when trained on new data. The proposed framework specifically addresses this by decoupling encoder training from decoder adaptation.
- Anchor-Aided Training Strategy
- A novel two-stage training method. First, a shared encoder and symmetric decoder are trained together to learn robust semantic representations. Then, the optimized encoder is frozen to train user-specific decoders independently, ensuring stable learning without interference.
Terminology
Summary
Semantic communication (SemCom) can actively embed the semantic meaning of the data into the transmission process, while eliminating statistical redundancy to preserve bandwidth resources for other users Most existing works only consider one transmitter and one receiver, which limits their ability to address the diversity in users’ models and capabilities Therefore, in this paper, we propose a multi-user semantic communication system where each user is equipped with a distinct DL-based joint source-channel decoder architecture, reflecting the diversity in computing capacity The challenging issue with the proposed system is the catastrophic forgetting property of neural networks, where the DL-based encoder fails to encode the data for the previous user when being trained with a new user. p. 0
System Architecture and Heterogeneity
The proposed system involves a downlink transmission from a base station (BS) to multiple users, each equipped with a distinct DL-based decoder architecture tailored to its computational capability This cross-architecture poses a critical challenge for training DL models in the network: the shared BS encoder received different updated directions from clients in the network, which leads to the catastrophic forgetting. The system considers users architecturally heterogeneous, reflecting the asynchronous and resource-constrained nature of practical devices in a real-world network. p. 1
Anchor-Aided Training Strategy
To address the gradient conflict problem, a novel two-stage training framework is proposed that decouples encoder optimization from decoder adaptation In the first stage, the encoder is trained together with a decoder that shares a symmetric architecture, which promotes stable learning and allows the encoder to fully capture semantic representations. The parameters of the optimized encoder are then frozen and used to train decoders for various users, aligning them with the encoder outputs. p. 0
Theoretical Foundation of Forgetting
Theorem 1 shows that when gradients from different users are negatively aligned, the loss of user j increases after the update, indicating conflicting update directions and ultimately resulting in catastrophic forgetting Theorem 2 demonstrates that after the self-reflective training stage, the optimization problems for each user decoder become independent, meaning no gradient interference occurs across users.
Performance Validation
Extensive simulations over both additive white Gaussian noise (AWGN) and Rayleigh fading channels demonstrate the effectiveness of the proposed framework compared to conventional iterative and joint training schemes In general, our framework achieves the highest performance across all types of decoder users and a wide range of channel qualities. The high performance is achieved because the symmetric has successfully guided the BS’s encoder to extract the data meaning and compress it in a meaningful way. p. 6
Conclusion
In this paper, we propose a downlink communication framework from a single base station to multiple users using semantic communication Our proposed framework achieves higher performance compared to existing training schemes across almost all levels of channel noise in both channel models. The symmetric has successfully guided the BS’s encoder to extract the data meaning and compress it in a meaningful way. p. 6
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Anchor-aided two-stage training framework allows for decoupling encoder optimization from decoder adaptation to stabilize learned representations while preserving flexibility for arbitrary user numbers. This enables efficient adaptation across users without degrading the representations learned by the encoder, as stated in
the parameters of the optimized encoder are then frozen and used to train decoders for various users, aligning them with the encoder outputs.
-
The framework prevents catastrophic forgetting by training a single batch of BS encoder parameters with a symmetric decoder first, which is described as
a form of self-reflective learning applied to the encoder.
This stage ensuresstable convergence during training and enhances the BS encoder’s semantic extraction and compression.
-
The system can handle architecturally diverse users by equipping each user with a distinct DL architecture, reflecting
the asynchronous nature of device resources in real-world environments.
This allows for different groups of clients to have varying layer counts, simulating differences in computing capacity. -
The proposed framework achieves higher performance across all types of decoder users and a wide range of channel qualities compared to iterative and simultaneous training cases. Specifically, it demonstrates
the highest performance across all types of decoder users and a wide range of channel qualities.
Sources
Related papers
- The Gaussian Multiuser Channel Capacity Region
- The mathematical landscape of partial information decomposition: A comprehensive review of properties and measures
- Electromagnetic Signal and Information Theory: A Continuous-Aperture Array Perspective
- Analytically Characterized Optimal Power Control for Signal-Level-Integrated Sensing, Computing and Communication in Federated Learning
- Information Spectrum Methods for Multiple-Access Channels--Part I: epsilon-Capacity Regions in the Mixed Gaussian Case
- Seed Recycling: Explicit Codes for Mixed Synchronization Errors Matching the Leading Term of the Gilbert--Varshamov Bound