Fig. 5 | Nature Communications

Fig. 5

From: Model-free inference of direct network interactions from nonlinear collective dynamics

Fig. 5

Learning curves with respect to the number of iterations l reveal appropriate expansions. af Learning curves for basis functions shown in Table 1; indices in Table 1 equal panel identification. The training and validation sets are composed by the 60 and 40% of recordings, respectively. For all basis functions ad chosen from an appropriate class (here: true pairwise interactions), the curve exhibits an L-shape with a clear plateau starting at the correct number of incoming interactions (vertical dashed line), above adding more interactions reduces the fitting costs only weakly or not at all. If basis functions are chosen from an inappropriate class (e, f), the cost functions decrease only mildly and do not exhibit L-shape

Back to article page