Summary: weekly mtg 20161019 (Byron, Steve, Matt, me) - mobeets/nullSpaceControl GitHub Wiki
Discussed
- SSS: looks good; use 8 directions to include nelson
- need consistent/objective metric for behavioral asymptote
- find paper by Diedrichsen
for next week:
- fit with new inclusion metric: does it change results?
- writing: fit more reference-material perspectives into Intro
- writing: add in SSS to the end: "because most activity might be null in both mappings, we don't want our results dominated by this activity. so we focus on this subspace"
- look into submitting to cosyne (finish wed before friday before deadline for sharing with scabby)
progress metric:
- progress under last 100 trials of intuitive through perturbation
- (non-overlapping?) variance in perturbation
- threshold variance (some multiple of the variance under intuitive?)
- pick section with best mean, compare it (t-test) to intuitive activity, then
To discuss
SSS
Nelson's data
Minimum energy as factors, with noise
Spike version:
- Find spikes with minimum norm such that latent kinematics are satisfied.
- Add gaussian spiking noise using spikeCountStd.
- Convert to latents (will satisfy latent row space values by design.)
Latent version:
- Find latents meeting cursor velocity while minimizing norm of null space value.
- Subject to min and max total factor activity (using held out data).
- Spiking noise: generate under factor analysis model (i.e., gaussian noise), then infer latents.
(And then, could set to nan anything that now has invalid spiking values, but I don't.)