GBPNet
A neural decoder architecture for linear code decoding. While preserving the message-passing symmetry of the Tanner graph, it introduces learnable components such as element-wise message scaling and cascaded variable node updates to improve decoding accuracy.
Timeline (last 90 days)
GBPNet (Generalized Belief Propagation Network) is a neural decoder architecture that improves the learning capability of the decoder while preserving the symmetry of message passing.
GBPNet adopts a cascaded variable node update scheme, so that each variable node receives and integrates more information within a single iteration, rather than strictly alternating by layer.
GBPNet introduces learnable element-wise message scaling to adaptively adjust the magnitude of the messages passed.
GBPNet generalizes message scaling into an element-wise, learnable scaling vector, so that each message dimension obtains an independent correction magnitude without changing the topology of message passing.
The paper proposes two training strategies, progressive training and high-SNR training, which do not change the network structure and only adjust the training procedure.
Experiments on multiple linear codes show that GBPNet achieves a lower bit error rate (BER) than traditional belief propagation and the evaluated state-of-the-art neural decoders.
The paper's evaluation scope is limited to the selected linear codes and experimental conditions, and whether GBPNet fully avoids the computational overhead and generalization issues commonly faced by neural decoders still requires further verification.
All Facts (7)
GBPNet (Generalized Belief Propagation Network) is a neural decoder architecture that improves the learning capability of the decoder while preserving the symmetry of message passing.
50%UnverifiedGBPNet adopts a cascaded variable node update scheme, so that each variable node receives and integrates more information within a single iteration, rather than strictly alternating by layer.
50%UnverifiedGBPNet introduces learnable element-wise message scaling to adaptively adjust the magnitude of the messages passed.
50%UnverifiedGBPNet generalizes message scaling into an element-wise, learnable scaling vector, so that each message dimension obtains an independent correction magnitude without changing the topology of message passing.
50%UnverifiedThe paper proposes two training strategies, progressive training and high-SNR training, which do not change the network structure and only adjust the training procedure.
50%UnverifiedExperiments on multiple linear codes show that GBPNet achieves a lower bit error rate (BER) than traditional belief propagation and the evaluated state-of-the-art neural decoders.
50%UnverifiedThe paper's evaluation scope is limited to the selected linear codes and experimental conditions, and whether GBPNet fully avoids the computational overhead and generalization issues commonly faced by neural decoders still requires further verification.
50%