Fisher information of the IMNN during training. Dark (light) blue curves show the Fisher information for the training (validation) set of the IMNN trained at θ_fid. The violin plots show the distribution of the Fisher information over the full prior volume excluding θ_fid, P(θ), at every interval of the 10k training epochs. Black dots and error bars inside each violin denote the median and 68% CLs of the distributions. The divergence between the Fisher information computed at θ_fid and its PDF over the remaining prior volume clearly indicates that the IMNN is over-specializing to the specific parameter combination, θ_fid, used in the training.