Table 3 Hyperparameter choices justification.
Hyperparameter | Value | Justification |
|---|---|---|
Learning rate | Default / 0.001 (assumed) | A moderate learning rate ensures stable convergence without overshooting minima |
Number of Conv layers | 5 Conv + 5 Pool | Deeper architecture enables learning complex features; each Conv-Pool pair extracts hierarchical patterns |
Filter sizes | 3 × 3 | Standard kernel size for spatial locality while reducing parameters compared to larger kernels |
Number of filters | 32 → 64 | Increasing depth allows richer feature extraction while keeping the model size manageable |
Activation function | ReLU (in Conv), Softmax (in Output) | ReLU avoids vanishing gradients; Softmax is ideal for multi-class classification |
Batch size | 32 | A common choice balancing GPU memory use and training stability |
Optimizer | Assumed Adam | Adaptive learning rates make it suitable for most CNN-based classification tasks |
