Table 3 Tested values of the hyper-parameters and their best values for the pan-cancer prediction tasks
Hyper-parameters | Tested range | Microarray | TCGA cancer pred. | TCGA type pred. |
|---|---|---|---|---|
Nb layers | [1, 20] | 4 | 5 | 4 |
Nb neurons | [20, 2000] | 600-600-600-60 | 500-500-500-500-50 | 700-700-700-50 |
Batch norm. | Yes/no | No | Yes | Yes |
Dropout | [0, 0.8] | 0 | 0 | 0 |
Optimizer | SGD/RMSprop/ADAM | SGD | SGD | SGD |
Learning rate | \([1^{-6},1]\) | \(10^{-3}\) | \(10^{-3}\) | \(5.10^{-2}\) |
L1 regularization | \([0,1^{-2}]\) | 0 | 0 | 0 |
L2 regularization | \([0,1^{-2}]\) | 0 | 0 | 0 |
Batch size | [8, 1024] | 8 | 32 | 8 |