twpca
twpca copied to clipboard
New loss functions
Users should have the ability to specify different loss functions. A few easy ones:
- [ ] Logistic loss (spiking data where there is at most 1 spike per bin)
- [ ] Poisson loss (spiking data with multiple spikes per bin)
- [ ] Gamma loss (for calcium imaging data)
In the case of Logistic/Poisson loss functions, this would remove the need to smooth the data as a preprocessing step (see https://github.com/ganguli-lab/twpca/issues/2), but would necessitate adding regularization for smoothness on the temporal factors. This could also help with https://github.com/ganguli-lab/twpca/issues/3 - e.g. the reconstruction can be interpreted as a probability of spiking in the logistic case.