Appendix 1—figure 2. Effect of changing conductor smoothing kernels in the plasticity rule.

(A) Matrix showing learning accuracy when using different timescales for the student plasticity rule. Each entry in the heatmap shows the average rendition error after 1000 learning steps when pairing a tutor with timescale ${\tau }_{\text{tutor}}$ with a non-matched student. Here the kernels are exponential, with timescales ${\tau }_1=20\mathrm{ms}$, ${\tau }_2=10\mathrm{ms}$. (B) Evolution of motor error with learning using kernels $\sim e^{-t/\tau }$ and $\sim t{e}^{-t/\tau }$, instead of the two exponentials used in the main text. The tutor signal is as before, Equation (3). The inset shows the final output for the trained model, for one of the two output channels. Learning is as effective and fast as before.

DOI: http://dx.doi.org/10.7554/eLife.20944.017