Fig. 3

Code length savings with synthetic data. The code length savings l as a function of the threshold \(\tau\) for different sample sizes and planted covariance \(c=0.3\). With enough samples we get a positive code length saving in the test network, but with fewer samples there is no signal of the modules in the train network. The code length savings are negative when there are spurious modules, or zero when there is only one module. Calculations are averaged over ten runs and the shaded areas show the standard deviation