Hyperparameter selection via held-out test sets.

Pattern noise ϵ_p, the probability of a deletion p_d and the Jensen-Shannon divergence threshold for removing similar bouts, J_thr, were set by computing the free energy per bout on a randomly split held-out dataset of approximately 18000 bouts (training dataset size was 70000). The free energy per bout for each tested parameter triplet is computed by averaging over 18 trials (with a separate random split for each trial). The red square shows the parameters (ϵ_p = 0.1, p_d = 0.2, J_thr = 0.15) chosen for the rest of our analysis.

(EPS)