Meta-Learning Enhanced BiLSTM Autoencoder with Channel-adaptive Quantization for Robust One-Bit Error Correction Coding

We propose a meta-learning-enhanced BiLSTM autoencoder architecture for robust one-bit error correction coding, designed to dynamically adapt to diverse channel conditions without requiring explicit retraining. The proposed method fuses a channel-aware meta-discriminator into an adversarial training framework, allowing the system to generalize across Rician, Rayleigh, and AWGN channels by adapting its decision boundaries based on temporal signal statistics. The meta-discriminator, realized as a lightweight Transformer-encoder with cross-attention, computes channel-specific embeddings from the received signal, which modulate the adversarial loss and guide the reconstruction process. Furthermore, the BiLSTM encoder-decoder utilizes bidirectional layers with residual connections to capture long-range dependencies, while a learnable one-bit quantizer with adaptive thresholds ensures efficient signal representation. The training objective combines reconstruction loss, adversarial loss, and a meta-regularization term, which stabilizes updates and refines adaptation. The meta-discriminator performs real-time parameter adjustments using a single gradient step during inference to make the system resilient to unseen channel impairments. The experiments demonstrate significant improvements in BER and MSE across various fading channels and data sizes. The Rician channel exhibits the lowest values of BER and MSE of 0.032 and 0.031, respectively, when considering a data size of 2500 symbols. The proposed work shows its dual capability to learn error-correcting codes through BiLSTMs, apart from exploiting meta-learning for channel adaptation.