Semantic Communication PHY

Unlike traditional systems that separate source and channel coding, Semantic PHY uses a single neural network to jointly encode the source data directly into channel symbols. This eliminates the 'cliff effect' of digital systems, allowing for graceful degradation: as channel quality worsens, the reconstruction becomes slightly less detailed rather than failing completely. Architectures like DeepJSCC map images directly to complex-valued I/Q samples, bypassing bits entirely.

The physical layer is optimized not for generic data reconstruction but for a specific downstream task at the receiver. For example, in a surveillance scenario, the transmitter might only send features relevant for object classification rather than pixel-perfect video.

• Receiver-defined semantics: The receiver's task dictates what information is 'meaningful'.
• Feature compression: Only the latent features necessary for inference are transmitted.
• Bandwidth efficiency: Dramatically reduces the required data rate by ignoring task-irrelevant information.

Classical transforms like the Fast Fourier Transform (FFT) or Discrete Cosine Transform (DCT) are linear and fixed. Semantic PHY employs learned, non-linear transforms via deep neural networks. The encoder learns a complex, high-dimensional mapping of the source signal into a latent semantic space optimized for transmission over a stochastic channel. This allows the system to discover efficient representations that no hand-crafted transform can achieve.

The entire communication chain—transmitter, channel model, and receiver—is implemented as a single differentiable computational graph. This allows for end-to-end training using backpropagation. The channel must be modeled as a differentiable layer (e.g., additive noise, stochastic fading). This joint optimization ensures the encoder learns representations that are both semantically compact and robust to the specific channel impairments, a feat impossible with modular, block-based design.

The training objective shifts from minimizing bit-error rate (BER) or symbol-error rate (SER) to maximizing the mutual information between the transmitted semantic latent and the receiver's task output. This is often implemented via a variational information bottleneck (VIB) or contrastive learning objective.

• Information Bottleneck Principle: The encoder is forced to compress the source into a minimal sufficient statistic for the task.
• Loss function: A weighted sum of rate (latent entropy) and distortion (task performance).

Semantic PHY systems can dynamically adapt their encoding strategy based on real-time Channel State Information (CSI) feedback. An attention mechanism or a hyper-network can condition the encoder's weights on the current signal-to-noise ratio (SNR) or fading profile. This enables the system to allocate more 'semantic bandwidth' to critical features when the channel is good and gracefully reduce fidelity when it is poor, all within a single, unified model.