improvement on minimum/baseline, where it uses true data?

[mobeets]

aaron's idea, to parallel what we'll have with uncontrolled after #268 is this possible?

[mobeets]

instead of solving for a spike, pick from intuitive sessions the point in spike null space that would bring you closest to minimum/baseline spikes

basically, this is exactly the same as what unconstrained does, except instead of sampling from the null space activity produced in the intuitive session, it picks the point that brings the total activity closest to baseline/minimum spikes. (also it works in spike space and not latent space.)

[mobeets]

[mobeets]

(also, it's definitely a lot better in spike space than fitting it in latent space)

[mobeets]

oh shit

[mobeets]

i think it's doing better here because this hyp does better when there's fewer points in the intuitive session, for whatever reason.

but it also does better with learning, which is negatively correlated with intuitive trial count.

[mobeets]

cloud, by comparison, does better with more intuitive trials

[mobeets]

[mobeets]

strangely enough, they're kinda indistinguishable from one another, baseline and minimum:

[mobeets]

after adding more dates, and only keeping those with session-wide behavioral asymptote

[mobeets]

[mobeets]

[mobeets]

but if we only analyze sessions where Lrn > 0.7, cloud is still better:

[mobeets]

wow, well i made a major mistake: this new minimum stuff was using all perturbation data by mistake! i.e., never using intuitive, only cheating!

that makes so much more sense. sheesh.