Supplementary Figure 2: Comparison of default parameters and parameter variations during learning.

The results of various alternative parameter settings are plotted against the reference labels produced using default settings in the training data set (n = 566,223; see Supplementary Table 2 for summary statistics). Most alternative settings produce predictions that are highly correlated with the reference. ELOG1.1, the E step uses a logarithmic upper bound with base of 1.1 (default = 2); ELOG10, the logarithm base is 10; ECONSTANT3, the E step uses a constant upper bound set to 3 (the default upper bound scales logarithmically in base 2); ECONSTANT10, constant upper bound of 10; ECONSTANT20, constant upper bound of 20; EFLOOR0, the E-step lower bound is set to 0 (default = 1); EFLOOR5, the lower bound is set to 5; E3to10, the E step uses lower and upper bounds of 3 and 10 for all samples; ESTEP0.8, the E-step sliding bound is calculated as 80% of current belief (default = 90%); ESTEP0.95, the sliding bound is calculated as 95% of current belief; MCV2, the M step uses 2-fold cross-validation (default = 5); LOGISTIC, the M step uses logistic regression (default is a tuned neural network); NODES6, the M step uses a neural network with only 6 hidden nodes (default is tuned, can use more than 6 nodes); DECAY0.1, the M step uses a neural network with weight decay of 0.1 (default is tuned, can use less stringent decay); DECAY0.1; NODES6, the M step enforces use of a simpler neural network than default settings require.