Previously, it was inflated at the lowest value. In an extreme case (where there are few target cells), it caused the FR to be zero for half of the population. (the figures below illustrates the fix)
The OSI loss is modified to include an offset that is proportional to the firing rates. (
With the previous implementation, the OSI loss is minimized at zero spikes (and fight against the rate-loss). With the new implementation, pure increase in firing rates is not penalized.
The neuron IDs are filtered before making the target firing rates. (it needs to be done before sampling the target firing rates)
The voltage loss also uses core_masks
Fixed the incorrect sampling of the target firing rates.