Figure 8.

Relationship of MF input to learned output influences how GCL supports learning. A, top: schematic of model task, using recorded EMGs as an input to the model GCL to predict kinematics. Bottom: MSE of model as a function of threshold when using EMG alone (MFs; blue) or GCL (red) as input to model P-cell. At a range of thresholds, P-cells that receive GCL input outperform P-cells receiving MFs alone. B: example of learned kinematic position after training for MF alone (blue line) and GCL network (red) showing good metric fit by the GCL model. C: plot showing the strength of different GCL population statistical features driving learning that vary as a function of how well MFs intrinsically support learned P-cell output (MF alone MSE). When MFs are already excellent predictors, information retention (variance retained) has a high regression slope (RIDGE regression method). When MFs are poorer intrinsic predictors, the number of explanatory PCs (a pattern separation metric) emerges as a stronger driver of learning performance. Goodness of fit (R²) was between 0.83 and 0.95 across all MF- alone MSEs used. GCL, granule cell layer; MF, mossy fiber; P-cell; Purkinje cells.