Table 3 The DNN estimates degradation parameters, and the layers used in the network.
Component/hyperparameter | Type and description | Value |
|---|---|---|
Input layer | Layer | FeatureInputLayer(size(4 for Depression − 2 for Anxiety)) |
Hidden layer 1 | Layer sequence | FullyConnectedLayer(64) → batchNormalizationLayer → reluLayer → dropoutLayer(0.3) |
Hidden layer 2 | Layer sequence | FullyConnectedLayer(32) → batchNormalizationLayer → reluLayer → dropoutLayer(0.3) |
Hidden layer 3 | Layer sequence | FullyConnectedLayer(16) → reluLayer |
Hidden layer 4 | Layer sequence | FullyConnectedLayer(8) → reluLayer |
Output layer | Layer sequence | FullyConnectedLayer(4 for Depression − 2 for Anxiety) → softmaxLayer → classificationLayer |
Optimizer | Training option: adaptive moment estimation optimizer used for weight updates | ‘Adam’ |
Max epochs | Training option: maximum number of training epochs | 100 |
Minibatch size | Training option: number of samples per training mini-batch | 16 |
Initial learning rate | Training option: starting value for the learning rate | 1e-3 |
Shuffle | Training option: shuffles data at the beginning of each epoch | ‘Every-epoch’ |
Validation patience | Training option: early stopping patience (stop if no improvement after 5 validations) | 5 |
Cross-validation folds | Evaluation scheme number of folds used for cross-validation | 10 |
