Channel Autoencoder

The foundational architecture where a transmitter and receiver are implemented as a single deep neural network. Unlike a standard autoencoder that reconstructs its input, this model learns a mapping from source bits to decoded bits through a stochastic channel layer. The joint optimization replaces the entire traditional physical layer, learning a robust, data-driven communication scheme without explicit modulation or coding blocks.

A critical enabler for gradient-based training. This is a mathematical or neural surrogate of the physical channel that allows backpropagation from the receiver loss to the transmitter weights.

• Stochastic Perturbation: Models additive white Gaussian noise (AWGN) with a reparameterization trick.
• Neural Surrogate: A pre-trained GAN or variational autoencoder (VAE) that mimics complex non-linear hardware impairments.
• Generative Adversarial Network (GAN): Used to model intractable channel effects like power amplifier non-linearity.

An extension of the channel autoencoder that bypasses separate source and channel coding. A single neural network directly maps raw source data—such as images or sensor readings—to channel symbols. This approach is optimized for bandwidth-constrained regimes, achieving graceful degradation where reconstruction quality improves smoothly with channel quality, unlike the cliff effect in digital systems.

A method where the autoencoder optimizes the positions and probabilities of points in the I/Q plane to maximize throughput for a specific channel model.

• Geometric Shaping: Moves points to non-uniform positions.
• Probabilistic Shaping: Assigns non-uniform probabilities to fixed points.
• Performance: Approaches the Shannon capacity limit by transmitting low-energy symbols more frequently, outperforming fixed QAM schemes on non-linear or non-Gaussian channels.

A transceiver designed to operate without explicit channel state information (CSI) . The neural network learns robust representations that are invariant to unknown channel phase and amplitude variations. This enables blind detection in high-mobility scenarios where pilot-based estimation is infeasible, learning to encode information in the magnitude or differential phase of the signal.

A paradigm shift from bit-exact transmission to goal-effective communication. The joint source-channel encoder is optimized to transmit only the semantic information relevant to a specific inference task at the receiver, such as image classification or object detection. This dramatically reduces bandwidth by discarding task-irrelevant data, a core principle for 6G semantic communication.