the issue isn't predicting latents, it's evaluating in terms of latents.
so for all the hyps that predict latents, we can then just generate spikes from them under the FA model...and then evaluate in terms of spikes.
min/bas are the only hyps that predict spikes on their way to predicting latents. (though they could predict latents directly with a simple toggle.) the other hyps all directly predict latents.
Proposal: Make Supp. Fig 6: scores when evaluating the same exact predictions in spike space, after converting predicted latents to spikes.
Things done:
Output-potent vals same with spikes or latents? No.
Make a cloud hyp using spikes, then project to latents just as min/bas do. Is that good? Gets about 50% histError per monkey; though it has to reject a lot of points due to them being invalid spike counts.
Fit and eval all hyps in spike space might be better supp. fig than the "fits in reverse" thing. histErr here, after doing PCA on the observed spike null activity and keeping the top 5 cols to score, we get cloud with histErrs of 17, 12, and 22% for each monkey. didn't run any other hyps though.
Finally, I just convert predicted latents to spikes, then evaluate in spike space (n.b. I also first remove any row-space info from latents because FA isn't an orthogonal projection. See attached plot for results. To be fair to min/bas I should just use the raw spikes they predict though
Note:
Proposal: Make Supp. Fig 6: scores when evaluating the same exact predictions in spike space, after converting predicted latents to spikes.
Things done: