Table 1
Hyperparameter of proposed approach.
| Hyperparameter | Value(s) | Justification |
|---|---|---|
| Epochs | 20 epochs with early stop | Ensures sufficient training; early stopping avoids overfitting and halts training once performance plateaus. |
| Learning Rate | Task 1: 0.00002 Task 2: 0.0001 to 0.000025 Task 3: 0.0002 to 0.0001 |
Stable convergence for Task 1; Task 2 and Task 3 uses a dynamic range to optimize learning and prevent destabilization. |
| EWC Lambda | Task 1: Not used Task 2: 14863.87 to 10308.57 Task 3: 14863.87 to 10308.5 |
Used only in Task 2 and Task 3 to prevent catastrophic forgetting; higher values indicate strong regularization of previous knowledge. |
| EWC Penalty (γ) | Task 1: Not used Task 2: 743.19 to 515.43 Task 3: 743.19 to 515.43 |
Applies regularization strength in EWC; helps balance plasticity and stability in Task 2. |
| Dropout Rate | 0.25 | Helps prevent overfitting by randomly deactivating neurons during training. |
| Batch Size | 16 | A small batch size allows for more updates per epoch and can improve generalization. |
| Training Size of Data | 80% | A large portion of data ensures effective model learning. |
| Testing Size of Data | 20% | Reserved to evaluate model performance and detect overfitting. |
| Loss Function | Cross-Entropy Loss | Commonly used for classification tasks; measures the performance of output probabilities. |
| Optimizer | ADAMW | Combines the benefits of Adam with weight decay for better generalization. |
| Activation Function | ReLU | Efficient and widely used; introduces non-linearity and avoids vanishing gradients. |
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.