Neural Channel Coding

Unlike classical codes designed analytically, neural channel codes are trained using stochastic gradient descent through a differentiable channel model. The encoder and decoder are jointly optimized as a single autoencoder, with backpropagation flowing from the decoder's bit-error loss through the channel layer to the encoder's weights. This allows the system to learn non-linear mappings that adapt to specific impairments like amplifier saturation, phase noise, or interference patterns that algebraic codes cannot model.

Neural encoders learn a continuous, high-dimensional embedding of information bits into channel symbols, rather than selecting from a fixed, discrete codebook. The decoder learns a corresponding non-linear decision boundary in the received signal space. This geometric approach enables:

• Shaping gain: Learned constellations naturally approximate a Gaussian distribution
• Coding gain: The embedding maximizes Euclidean distance between codewords under the specific channel distribution
• Joint optimization: Modulation and coding are not separate stages but a single learned mapping

Rather than replacing classical decoders entirely, model-based architectures integrate known algorithmic structure as non-trainable layers. Examples include:

• Neural Belief Propagation: Graph neural networks operating on the Tanner graph of an LDPC code, learning optimized message-passing weights
• ViterbiNet: A neural network that replaces the hand-crafted branch metric in the Viterbi algorithm, enabling decoding over channels with unknown memory
• Hypernetwork Decoders: A meta-network that generates decoder weights conditioned on estimated channel state, enabling single-model adaptation across varying conditions

Classical codes assume additive white Gaussian noise (AWGN) and linear channels. Neural codes excel where these assumptions break down:

• Non-linear hardware impairments: Power amplifier distortion, I/Q imbalance, phase noise
• Compound channels: Simultaneous fading, interference, and jamming
• Finite blocklength regimes: Short-packet communication where asymptotic information-theoretic bounds are loose
• Unknown channel models: The network learns the effective channel distribution from data without explicit mathematical characterization

A single neural codec can be trained to support multiple code rates and tasks without redesign. Techniques include:

• Rate-compatible architectures: A mother code where puncturing patterns are learned, not predefined
• Conditional encoders/decoders: An additional input specifies the desired rate, modulation order, or target block error rate
• Joint source-channel coding: The same network maps raw source data (images, sensor readings) directly to channel symbols, adapting the coding rate implicitly based on source entropy and channel quality
• Incremental redundancy: The decoder requests additional parity symbols only when decoding fails, with the encoder generating them on-demand

Training neural channel codes presents unique difficulties addressed by specific techniques:

• Gradient estimation through stochastic channels: The channel layer is non-differentiable for discrete outputs. Solutions include the Gumbel-Softmax reparameterization or training with a differentiable channel surrogate (e.g., a mixture of Gaussians approximating the true channel)
• Overfitting to training SNR: A model trained at a single signal-to-noise ratio may fail at others. Multi-SNR training samples a range during optimization
• Curse of dimensionality for long blocks: For large blocklengths, Transformer-based codecs with self-attention capture long-range dependencies better than recurrent architectures
• Decoder complexity: Neural decoders must match the latency constraints of hardware. Knowledge distillation trains a smaller student network to mimic a larger teacher decoder

A neural network architecture that jointly optimizes a transmitter and receiver as a single deep learning model. It replaces traditional block-based communication algorithms with a learned, data-driven mapping from source bits to decoded bits. The entire system is trained over a differentiable channel model or directly over-the-air, allowing the encoder and decoder to co-adapt to hardware impairments and non-linear channel effects that classical codes cannot easily model.

A neural decoder that operates directly on the Tanner graph structure of a linear block code. Using graph convolutions and learned message passing, it mimics and improves upon belief propagation algorithms. Key advantages include:

• Faster convergence than classical BP
• Robustness to correlated noise
• Direct learning from data without requiring explicit channel statistics This approach bridges classical algebraic code structure with data-driven optimization.

A sequence-to-sequence communication model based on the self-attention mechanism. Unlike recurrent neural network-based decoders, the transformer codec processes long temporal dependencies in coded bit streams in parallel. It excels at iterative decoding tasks where long-range interactions between codeword bits are critical, learning to attend to relevant parts of the received sequence without the vanishing gradient limitations of RNNs.

A mathematical or neural surrogate model of a physical communication channel that allows gradients to backpropagate from the receiver loss to the transmitter parameters. This is the critical enabler for gradient-based end-to-end optimization of the entire transceiver. Without a differentiable channel, the transmitter cannot receive a learning signal. Common approaches include:

• GAN-based channel approximations
• Mixture density networks
• Direct over-the-air training with stochastic gradient estimation

A family of learned forward error correction schemes where a neural encoder maps messages to codewords and a neural decoder recovers them under noise. Unlike classical algebraic codes with fixed mathematical structure, neural codes learn the optimal encoding manifold and decoding boundary for a specific channel distribution. Variants include neural block codes, convolutional codes, and turbo-like iterative architectures trained end-to-end.

A model-based deep learning receiver that integrates the Viterbi algorithm's structural prior with a neural network. It learns to decode sequences over channels with unknown, complex memory by replacing the hand-crafted branch metric calculation with a learned function. This hybrid approach preserves the algorithmic efficiency of dynamic programming while gaining the adaptability of neural networks to real-world channel impairments.