Figure 2
Figure 2 illustrates the algorithm used to select stable weight parameters from a trained neural network model. The input to the algorithm is the set of trained weight parameters w obtained from model M, and the output is the set of selected stable encoder weights W subscript E. The process begins by initializing a list S subscript G to store the average gradient values of each layer. For each weight node in the model, the gradient is computed. Then, for each layer in the model over the dataset O subscript k, the average gradient magnitude G subscript k is calculated and stored together with the corresponding layer index in S subscript G. The list S subscript G is subsequently sorted in ascending order according to the average gradient magnitude, allowing the layers with the smallest gradient values to be identified as the most stable. The first N layers with the lowest G subscript k values are selected. Finally, the weight parameters corresponding to the selected stable layers are extracted from the encoder and stored in the set W subscript E, which is returned as the output of the algorithm.Pseudocode for stable weights selection