Table 4 FFNN hyperparameter space explored with hg19 and hg38 data through Bayesian optimization. A. Architecture and learning hyperparameters of the fixed-FFNN; B. Architecture and hyperparameter space of the Bayesian-FFNN models trained on the hg19 dataset; C. Architecture and hyperparameter space of the Bayesian-FFNN models models trained on the hg38 dataset
Layers | Units | Activation |
|---|---|---|
A: fixed-FFNN | ||
Dense | 16 | ReLU |
Dense | 4 | ReLU |
Dense | 2 | ReLU |
Output | 1 | Sigmoid |
Learning parameters | ||
Learning rate | 0.5 | |
Learning rate decay | 0.1 | |
l2 regularizer | 0.0 | |
Batch size | 32 | |
Optimizer | SGD | |
Max no. of epochs | 64 | |
Layers | Hyperparameter space | Activation |
|---|---|---|
B: Bayesian-FFNN (hg19 dataset) | ||
No. of dense layers | {0, 1, 2, 3} | |
No. of units layer 1 | {256, 128, 64, 32, 16, 8, 4, 2} | ReLU |
No. of units layer 2 | {128, 64, 32, 16, 8, 4, 2} | ReLU |
No. of units layer 3 | {64, 32, 16, 8, 4, 2} | ReLU |
Output | 1 | Sigmoid |
Learning parameters | ||
Learning rate | [0.1, 0.5] | |
Learning rate decay | [0.01, 0.2] | |
l2 regularizer | [0, 0.1] | |
Batch size | [32, 256] | |
Optimizer | SGD | |
Max no. of epochs | [32, 1000] | |
C: Bayesian-FFNN (hg38 dataset) | ||
Groups | \(n=4\) | |
No. of hidden layers, composing the group | \(\{0, \ldots , 3\}\) | |
No. of units of the dense layer | \(\{0, \ldots , 256\}\) | ReLU |
Dropout | \([0, \ldots , 0.5]\) | |
Output | 1 | Sigmoid |
Learning parameters | ||
Learning rate | [0.1, 0.5] | |
Learning rate decay | [0.01, 0.2] | |
l2 regularizer | [0, 0.1] | |
Batch size | [32, 256] | |
Optimizer | SGD | |
Max no. of epochs | [32, 1000] | |