Summary of generalization errors for different training stimuli.

Left column denotes the five training sets employed: step current pulses only, ramp current pulses only, combined step and ramp current pulses and the two noise inputs. Models were generated by each one of these training sets and tested on four different generalization sets (top row): steps, ramps and type 1 and type 2 colored noise currents (see Materials and methods). Accuracy of generalization to step and ramp currents is measured as the average ± sd of all feature-based errors across all six features, for the acceptable models (lower values indicate greater accuracy). Accuracy of generalization to colored noise current injections is given by the precision of spike timing as quantified by the average gamma coincidence factor value (higher values are more accurate; see Materials and methods). Italicized text indicates within-stimulus generalization, regular indicates between stimulus generalization.