Deep Joint Source-Channel Coding

The foundational architecture for DJSCC, representing the transmitter and receiver as a single neural network. Unlike traditional systems with separate source and channel codecs, an end-to-end autoencoder learns a joint mapping from raw input data (e.g., pixels) directly to channel symbols and back to reconstructed data.

• Bottleneck layer represents the channel, forcing the network to learn robust, compressed representations.
• Trained with a single loss function, such as Mean Squared Error (MSE) or perceptual loss, rather than optimizing for bit error rate.
• Eliminates the cliff effect, where digital systems fail catastrophically below a threshold SNR, by gracefully degrading reconstruction quality.

A semantic communication paradigm where the goal is not to reconstruct the source perfectly, but to preserve only the information relevant for a specific downstream task. In DJSCC, this means the encoder strips away task-irrelevant data to save bandwidth.

• For an image classification task, the encoder may discard texture details while preserving object shape and edges.
• Jointly optimized with a task network at the receiver, creating a variational information bottleneck.
• Dramatically reduces the required data rate compared to transmitting a full, perceptually lossless image.

A deep learning framework that formalizes the trade-off between compression and task relevance. It learns a stochastic encoding that is maximally informative about the target task while being minimally informative about the input.

• The objective function maximizes mutual information between the encoding and the task label, minus a penalty on the mutual information between the encoding and the input.
• Naturally handles rate-distortion optimization in a task-aware manner.
• Provides a principled mathematical basis for designing DJSCC systems that go beyond pixel-wise distortion metrics.

A critical enabler for gradient-based training of DJSCC systems. Since a real physical channel is non-differentiable, a surrogate model must be used during training to allow backpropagation from the receiver loss to the transmitter weights.

• Can be a simple statistical model (e.g., AWGN, Rayleigh fading) implemented as a stochastic layer.
• For complex channels, a Generative Adversarial Network (GAN) can learn a differentiable surrogate from real measurements.
• The accuracy of this model directly impacts the real-world performance of the trained transceiver.

The broader category encompassing DJSCC, which shifts the objective from accurate bit transmission to the successful conveyance of meaning. Semantic communication systems are goal-oriented, transmitting only the semantic features necessary for the receiver's intended task.

• Moves beyond Shannon's technical level to the semantic and effectiveness levels of communication.
• Often uses joint source-channel coding as the underlying mechanism to map semantic features directly to the physical layer.
• Key for 6G applications where bandwidth is constrained but the value of information is high, such as autonomous driving sensor sharing.

A technique often integrated into DJSCC transmitters where the neural network learns the optimal geometric positions and probabilities of I/Q constellation points. This replaces fixed schemes like QAM with a data-driven modulation.

• The constellation is treated as a trainable parameter, optimized jointly with the encoder and decoder.
• Can learn non-uniform, non-grid shapes that are robust to specific channel impairments like phase noise.
• Combined with probabilistic shaping, it can approach the Shannon capacity limit by transmitting low-energy points more frequently.