Algorithm 1 RBG Neural Network

Input:

X ∈ ℝ^(T×d_in)    // Input sequence of length T with d_in features

n_layers = 6       // Total number of stacked layers

d_hidden        // Hidden state dimension

training_flag     // Boolean for train/inference mode

Output:

y ∈ ℝ^(T×d_out)    // Output sequence

1: // Initialization

2: for l ← 1 to n_layers do

3:    if l ≤ 2 then

4:      W_gru[l] ← Initialize GRU Weights(d_hidden)

5:    else

6:      W_bn[l] ← Initialize BatchNorm Params()

7:      W_gru[l] ← Initialize Bi-GRU Weights(d_hidden)

8:    end if

9: end for

10: // Forward pass

11: h ← X

12: for l ← 1 to n_layers do

13:    if l > 2 then

14:      h ← BatchNorm(h, W_bn[l], training_flag)

15:    end if

16:    h ← ReLU(h)  // Element-wise activation

17:    if l ≤ 2 then

18:      h ← GRU(h, W_gru[l])  // Unidirectional GRU

19:    else

20:      h ← BiGRU(h, W_gru[l])  // Bidirectional GRU

21:    end if

22: end for

23: y ← LinearProjection(h)  // Final output layer

24: return y