Table 3 Overview of multi-start settings per problem and Scenario
From: Current state and open problems in universal differential equations for systems biology
Scenario | # Starts | Hyperparameter | Values |
|---|---|---|---|
Glycolysis (per data set) | 10 000 | # Hidden layers | 1, 2, 3, 4 |
# Neurons/layer | 3, 5, 10 | ||
Activation function | tanh, relu, Gaussian rbf, swish | ||
Regularisation | None or λ ∈ [10−4, 10−1] | ||
ADAM learning rate | r ∈ [10−4, 10−1] | ||
Input normalisation | on, off | ||
STAT5 dimerisation Scenario 1-3 | 3 000 | # Hidden layers | 3 |
# Neurons/layer | 5 | ||
Activation function | tanh | ||
Regularisation | None or λ ∈ [10−3, 1] | ||
ADAM learning rate | r ∈ [10−4, 10−2] | ||
Input normalisation | on, off | ||
STAT5 dimerisation Scenario 4 | 30 000 | # Hidden layers | 1, 2, 3, 4 |
# Neurons/layer | 3, 5, 10 | ||
Activation function | tanh, relu, Gaussian rbf, swish | ||
Regularisation | None or λ ∈ [10−3, 103] | ||
ADAM learning rate | r ∈ [10−4, 10−1] | ||
Input normalisation | on, off |
