Table 2.
Weights on feature values to create fitness function.
| Feature | D1-SPN and D2-SPN | ArkyN and ProtoN |
|---|---|---|
| Baseline pre | 1 | 0 |
| Baseline post | 1 | 1 |
| Rectification | 0 | 2 |
| Falling curve | 1 | 1 |
| Voltage response | 1 | 1 |
| Latency | 1 | 0 |
| Spike time | 0 | 0.5 |
| Spike width | 1 | 1 |
| Spike height | 1 | 0.5 |
| Spike count | 1 | 1 |
| AHP depth | 1 | 1 |
| AHP curve | 4 | 1 |
| Histogram | 1 | 1 |
AHP curve was weighted higher for SPN to avoid the optimizer creating models with large, sharp AHPs. For the GPe models, spike height weight was reduced to avoid the optimizer producing an extreme mismatch in other features while trying to reduce spike height to the unusually small values observed experimentally. Spike time was reduced for both GPe and SPN models reflecting the high variability of this value between neurons of the same type. A weight of zero means to not use the feature, e.g., latency is not defined for a spontaneously spiking neuron. Features are further described in the online documentation (https://neurord.github.io/ajustador/features.html and https://neurord.github.io/ajustador/fitnesses.html). “histogram” is a root-mean-square of the difference between cumulative histograms of membrane potential-values in the two recordings.