Customized BiGRU Layer using Deep Learning Toolbox

BiGRU layer constructed based on Deep Learning Toolbox.

The Bidirectional Gated Recurrent Unit (BiGRU) layer consists of two independent GRU branches that process the same input sequence in forward and reverse orders. The forward GRU captures historical temporal information from past time steps, while the backward GRU extracts future contextual dependencies.

Benefiting from the reset gate and update gate inside GRU cells, BiGRU effectively mitigates the vanishing gradient problem of vanilla RNNs with fewer parameters than BiLSTM, balancing modeling capacity and training speed. This layer is widely used to extract bidirectional long-range dependencies for natural language understanding, time-series fault diagnosis and signal sequence modeling.