Turbo Autoencoder

Technical Explanation

Classical error-correcting codes (like LDPC and Turbo codes) were mathematically optimized under the assumption of ideal linear Gaussian channels. Real-world 6G hardware has non-linear power amplifiers, phase noise, and memory effects where classical codes are suboptimal. A Turbo Autoencoder trains deep convolutional neural networks end-to-end through physical channel hardware, discovering bespoke error-correcting codes that outperform classical codes on real-world silicon.

Key Functions

• End-to-end neural network channel coding outperforming classical codes on non-linear hardware
• Joint optimization of coding and modulation constellations via deep autoencoders
• Iterative neural belief propagation decoding matching Turbo code structural principles
• Autonomous adaptation to specific transceiver non-linearities and amplifier compression
• Hardware-accelerated neural inference executing on baseband tensor accelerators