Abstract: Recent works have shown that modern machine learning techniques can provide an alternative approach to the long-standing joint source-channel coding (JSCC) problem. Very promising initial results, superior to popular digital schemes that utilize separate source and channel codes, have been demonstrated for wireless image and video transmission using deep neural networks (DNNs). However, end-to-end training of such schemes requires a differentiable channel input representation; hence, prior works have assumed that any complex value can be transmitted over the channel. This can prevent the application of these codes in scenarios where the hardware or protocol can only admit certain sets of channel inputs, prescribed by a digital constellation. Herein, we propose DeepJSCC-Q, an end-to-end optimized JSCC solution for wireless image transmission using a finite channel input alphabet. We show that DeepJSCC-Q can achieve similar performance to prior works that allow any complex valued channel input, especially when high modulation orders are available, and that the performance asymptotically approaches that of unconstrained channel input as the modulation order increases. Importantly, DeepJSCC-Q preserves the graceful degradation of image quality in unpredictable channel conditions, a desirable property for deployment in mobile systems with rapidly changing channel conditions.
Introduction
In traditional wireless communication systems, the design follows the "separation principle," where source coding (compression) and channel coding (error correction) are handled by independent modules. This architecture, shown in Figure 1, has been the standard for decades. However, Deep Joint Source-Channel Coding (DeepJSCC) has recently emerged as a powerful alternative. Instead of compressing an image to bits and then adding error-correction codes, DeepJSCC uses an autoencoder to map images directly to complex-valued channel symbols.
A significant hurdle for DeepJSCC is that standard models assume the transmitter can send any complex number. In reality, commercial hardware—such as Wi-Fi routers and smartphones—is hard-coded to use specific sets of points called constellations (e.g., QAM or PSK). Because selecting a discrete point from a constellation is a non-differentiable operation, standard backpropagation cannot easily train a network to perform this task. DeepJSCC-Q bridges this gap between "analog-style" deep learning models and "digital-style" hardware constraints.
Methodology
The authors propose DeepJSCC-Q, a convolutional autoencoder architecture featuring a specialized quantization layer between the encoder and the noisy channel. The general framework for such a JSCC system, including precoding and equalization for wireless channels, is illustrated in Figure 2.
To make the system trainable, DeepJSCC-Q employs several key strategies:
- Soft-to-Hard Quantization: During the backward pass of training, the system uses a "soft" approximation where each symbol is represented as a softmax-weighted sum of all constellation points. During the forward pass (inference), it switches to "hard" quantization, snapping the output to the nearest real constellation point.
- Learned Constellations: Beyond standard grids like 16-QAM, the network can treat the constellation points as trainable parameters. This allows the network to "deform" the constellation to better suit natural image statistics and specific channel noise.
- KL-Divergence Regularization: A penalty term is added to the loss function to encourage the network to use all available constellation points uniformly, preventing "collapse" where only a few points are utilized.
Experimental Setup and Results
The models were trained on the ImageNet dataset (1.2 million images) and evaluated on the Kodak dataset, a standard benchmark for image compression.
Avoiding the Cliff Effect
Traditional digital systems suffer from a "cliff effect," where the image becomes completely un-decodable if channel noise exceeds a certain threshold. DeepJSCC-Q maintains "graceful degradation," where the image simply becomes blurrier as the signal weakens, similar to old analog television. This is achieved through emergent Unequal Error Protection (UEP), where the network learns to map critical image features to high-power constellation points.
Performance Benchmarks
- Superiority over Digital Baselines: DeepJSCC-Q consistently outperformed the combination of the BPG (Better Portable Graphics) codec and LDPC (Low-Density Parity-Check) codes, particularly in short-transmission (finite blocklength) regimes.
- Asymptotic Convergence: As the modulation order ($M$) increases (e.g., moving to 4096-QAM), DeepJSCC-Q's performance asymptotically approaches that of unconstrained, "analog" DeepJSCC.
- Fading Resilience: In slow-fading scenarios, learned constellations (e.g., L-64) outperformed standard 4096-QAM, demonstrating the advantage of optimizing constellation geometry for specific channel conditions.
Conclusion
DeepJSCC-Q provides a viable path for integrating deep learning-based communication into real-world hardware. While training requires careful parameter annealing and further work is needed to address hardware-specific issues like peak-to-average power ratio (PAPR), the results show that quantization constraints are not a dealbreaker for high-performance neural communication. Future research will explore integration into 5G and 6G standards where hardware-constrained modulators are already the norm.
Summary updated