Learning to Interfere in Non-Orthogonal Multiple-Access Joint Source-Channel Coding

The gist We introduce a machine learning (ML)-aided wireless image transmission method that merges compression and channel coding using a multi-view autoencoder, which allows the transmitters to use all the available channel resources simultaneously, resulting in a non-orthogonal multiple access (NOMA) scheme

How it works

The proposed method utilizes a novel multi-view autoencoder architecture for multi-user semantic image transmission using NOMA This architecture introduces user-specific projections, which enables shared encoder and decoder parameters across devices, crucial for scaling the solution to large wireless networks The system employs a novel superposition coding technique with user-specific projections, which allows network to scale to multiple users without significantly increasing the number of parameters and without changing the DeepJSCC architecture that is known to perform well

Key Method Components

1. Novel Scalable Multi-View Autoencoder Architecture: The paper presents a novel multi-view autoencoder architecture for multi-user semantic image transmission using NOMA This introduces user-specific projections, enabling shared encoder and decoder parameters across devices, crucial for scaling the solution to large wireless networks

2. Novel Progressive Fine-Tuning Strategy: A novel progressive fine-tuning strategy is introduced that doubles the number of users at each stage, building on any point-to-point DeepJSCC scheme This strategy maintains performance at the start of each stage thanks to orthogonalized user-specific projections

3. User-Specific Projections: The system introduces a novel superposition coding method with user-specific projections, which allows network to scale to multiple users without significantly increasing the number of parameters and without changing the DeepJSCC architecture that is known to perform well These projection matrices are initialized in a way that all of their rows are orthogonal, via QR factorization as detailed in Algorithm 2

4. Progressive Fine-Tuning Method: A novel progressive fine-tuning method is introduced by gradually doubling the number of users while sustaining the performance through orthogonal initialization of projection matrices This ensures that the normalized signals corresponding to different user groups are guaranteed to be orthogonal at this stage before any optimization

Performance and Analysis

The proposed scheme scales up to 16 users with only a 0.6% increase in the number of trainable parameters compared to a single-user model, significantly enhancing recovered image quality and outperforming existing NOMA-based methods over a wide range of datasets, metrics, and channel conditions The method outperforms TDMA-based DeepJSCC, NOMA alternatives, and separation-based methods with BPG, neural codecs, and LDPC across all signal-to-noise ratio (SNR) conditions for both independent and correlated source samples

Conclusion

Our DeepJSCC-PNOMA scheme outperforms DeepJSCC-NOMA and DeepJSCC-NOMA-CL over all SNR values Furthermore, it scales up to 16 users with only an extra 0.6% of trainable parameters at ρ¯ = 1/6 and 0.1% at ρ¯ = 1/12, demonstrating consistent performance gains The system is considered fair if the differences PSNRi−PSNRj are negligible for all i,j ∈[n] This consistency supports the generalization of our method to achieve high perceptual quality

Ablation and Fairness

The ablation study shows that parameter sharing not only boosts performance but also reduces training time Orthogonal initialization is the most effective component of our method, as it ensures that performance does not degrade due to randomness at the start of each fine-tuning step The results demonstrate that DeepJSCC-PNOMA consistently outperforms DeepJSCC-TDMA across all evaluated metrics, SNRs, and bandwidth ratios The method can be effectively used with a variable number of actively transmitting users, aligning with the comprehensiveness objective across most channel conditions

Technical Details

The per-user bandwidth ratio is defined as ρ≜ k CinWH channel symbols/pixel The goal is to maximize the average peak signal to noise ratio (PSNR), on an unseen target dataset defined as PSNR(x,xˆ) = 10log10 A squared 1 CinHW∥x−xˆ∥ squared 2 dB The overall number of network parameters is bounded by O Θ+Φ+n squared ρ¯ squared The training procedure involves jointly training the whole network with parameters Θ,Φ,Pi,P′i,∀i∈[n], using the loss function L = P(x1,…,xn) ∈ Dtrain Pn i=1 MSE(xi,xˆi), where m is the total number of elements in x The training durations mainly depend on the number of tuples T, per-user bandwidth ρ¯, training dataset size Dtrain, and the number of users n The memory requirement for evaluations is less than 3 GB of GPU memory

Appendix Details

The progressive fine-tuning procedure is illustrated in Figure 14 Algorithm 3 outlines the method used to construct the training and evaluation samples, which are detailed in Section IV-C The construction of training and evaluation samples involves generating a vector of random permutation of nT indices and matching them to the training data size N The datasets used include CIFAR10, TinyImagenet, Cityscapes, and Kodak We use the Pytorch framework for conducting all our deep learning experiments The LPIPS metric is employed to evaluate perceptual quality as it computes the similarity between the activations of two image patches for a pretrained neural network The MS-SSIM metric extends the SSIM by evaluating image quality at multiple scales and combining these measurements into a single score The LPIPS metric has become popular in image processing tasks like image super-resolution, GAN evaluation, and other applications where perceptual quality is crucial The MS-SSIM has been shown to perform better at representing human perception compared to PSNR The LPIPS metric is valued for its ability to better reflect human visual perception compared to traditional pixel-based metrics The results align with our discussion in Section V-C, despite being evaluated using different metrics

References

[2] L. Dai, B. Wang, Y. Yuan, S. Han, I. Chih-Lin, and Z. Wang “Non-orthogonal multiple access for 5G: Solutions, challenges, opportunities, and future research trends” IEEE Communications Magazine vol 53 no 9 pp 74–81 2015

[3] S. Verdu “Multiuser detection” Cambridge Univ 1998

[4] T. Cover, A. E. Gamal, and M. Salehi “Multiple access channels with arbitrarily correlated sources” IEEE Transactions on Information theory vol 26 no 6 pp 648–657 1980

[5] D. Gunduz, E. Erkip, A. Goldsmith, and H. V. Poor “Source and channel coding for correlated sources over a multiple access channel” IEEE Transactions on Information Theory vol 55 no 9 pp 3927–3944 2009

[6] B. Güler, D. Gündüz, and A. Yener “Lossy coding of correlated sources over a multiple access channel: Necessary conditions and separation results” IEEE Transactions on Information Theory vol 64 no 9 pp 6081–6097 2018

[7] A. Rezazadeh, J. Font-Segura, A. Martinez, and A. G. i Fàbregas “Joint source-channel coding for the multiple-access channel with correlated sources” in 2019 IEEE International Symposium on Information Theory ISIT IEEE 2019 pp 1317–1321

[8] E. Bourtsoulatze, D. B. Kurka, and D. Gündüz “Deep joint source-channel coding for wireless image transmission” IEEE Transactions on Cognitive Communications and Networking vol 5 no 3 pp 567–579 2019

[9] J. Xu, T.-Y. Tung, B. Ai, W. Chen, Y. Sun, and D. Gündüz “Deep joint source-channel coding for semantic communications” IEEE communications Magazine vol 61 no 11 pp 42–48 2023

[10] T.-Y. Tung and D. Gündüz “Deepwive: Deep-learning-aided wireless video transmission” IEEE Journal on Selected Areas in Communications vol 40 no 9 pp 2570–2583 2022

[11] T. Han, Q.

1. Scalable Multi-User Semantic Image Transmission: The system can transmit images from up to 16 users simultaneously over a multiple access channel using a novel multi-view autoencoder architecture with user-specific projections, achieving performance gains with only a 0.6% increase in trainable parameters compared to single-user models.

2. Progressive Fine-Tuning Strategy: The AI system can be trained efficiently for large numbers of users by employing a progressive fine-tuning algorithm that doubles the number of users at each stage, maintaining initial performance through orthogonalized user-specific projections and improving it through subsequent fine-tuning steps.

3. Interference Management via Orthogonal Initialization: The receiver can effectively distinguish between different users' inputs by initializing user-specific projection matrices in an orthogonal manner, which ensures that signals of different devices interfere with each other are managed without significant performance degradation during the initial stages of fine-tuning.

4. Robustness to Channel Variations: The deep joint source-channel coding (DeepJSCC) scheme is designed to be robust against channel variations, as the continuous-amplitude nature allows for graceful degradation with channel quality, maintaining high perceptual quality even under both AWGN and Rayleigh fading conditions.

5. Comprehensive Fairness Across Users: The system can maintain equitable image reconstruction quality across all users, even when only a subset of users transmit actively, satisfying the comprehensiveness objective where "PSNRP >= PSNR[n]" for any subset P.

We consider multiple transmitters aiming to communicate their source signals (e.g., images) over a multiple access channel (MAC). Conventional communication systems minimize interference by orthogonally allocating resources (time and/or bandwidth) among users, which limits their capacity. We introduce a machine learning (ML)-aided wireless image transmission method that merges compression and channel coding using a multi-view autoencoder, which allows the transmitters to use all the available channel resources simultaneously, resulting in a non-orthogonal multiple access (NOMA) scheme. The receiver must recover all the images from the received superposed signal, while also associating each image with its transmitter. Traditional ML models deal with individual samples, whereas our model allows signals from different users to interfere in order to leverage gains from NOMA under limited bandwidth and power constraints. We introduce a progressive fine-tuning algorithm that doubles the number of users at each iteration, maintaining initial performance with orthogonalized user-specific projections, which is then improved through fine-tuning steps. Remarkably, our method supports up to 64 users with potential for further scaling, with only a 2.4% increase in the number of trainable parameters compared to a single-user model, significantly enhancing recovered image quality and outperforming existing NOMA-based methods over a wide range of datasets, metrics, and channel conditions. Our approach paves the way for more efficient and robust multi-user communication systems, leveraging innovative ML components and strategies.