Fig. 2.

Predictive performance on naming scores. Model predictive performance is shown for: (a) linear support vector machines (light grey bars); (b) Gaussian processes with a rational quadratic kernel (mid-grey bars); and (c) the best of 16 inducers tried (dark grey bars). Models were trained on each of 7 data configurations: (i) lesion load only, L; (ii) restricted connectivity disruption, C(r); (iii) full connectivity disruption, C(f); (iv) lesion load plus restricted connectivity, LC(r); (v) lesion load plus full connectivity, LC(f); (vi) lesion load stacked with restricted connectivity, LsC(r); and, (vii) lesion load stacked with full connectivity, LsC(f). When produced using a linear support vector machine, there was a marginally significant benefit for the stacked model using lesion load and restricted connectivity (p = 0.04), and non-significant trend for the model which simply replaced lesion load with restricted connectivity (p = 0.07). No significant benefits were observed when predictions were made using either GPMR (all p > 0.1) or the best of 16 inducers (all p > 0.2). Numbers in each bar are prediction error distribution variances: all of the model comparisons are comparisons of these variances.