Architecture of the neural network used for the NDEs. The inputs are the model parameters and the summary statistics of a corresponding noised signal and the outputs allow us to evaluate the parametric likelihood function of the summary statistics (see main text). For each layer (blue regions), we give the number of neurons and the activation function. Then, N_c is the number of components of the GMM describing the likelihood, N_obs is the length of the vector of summary statistics, and $N_{\mathrm{off}}=\frac{(N_{\mathrm{obs}}+1)N_{\mathrm{obs}}}{2}$.