Summary: weekly mtg 20170419 (Byron, Steve, Aaron, Emily, me) - mobeets/nullSpaceControl GitHub Wiki
Progress gradient: at any point in time, per target, how should a neuron's firing rates evolve over time, based on these progress gradients?
- might be bad for quick, single day learning
- but maybe do it for factors, with changes in the learning rate that vary by facto...?
- subproblem of this is predicting which factors can learn
Tuning changes
- concern about latent/spike conversion
- subproblem: we don't know why PDs cluster (this will probably come up, so we'll probably have to find an explanation for it)
- Aaron says we could probably do some quick experiments
- might have to do with how we train the animals to use the BCI?
- maybe a long-term WMP that is crafted to pressure the PDs to uncluster
Predicting asymptotic performance
- need simulations maybe to see how well reassociation could do
- Emily mentions maybe doing long-term WMPs to see if eventually performance returns to baseline
Alan's running experiments, so is Nick; maybe sit in on some?
- get list of things I need to do from Emily
How is the monkey learning? Error-based, or reward-based?
Looking at washout: Is it consistent with both mappings?
- Seems quick to at least take a stab at it
- Byron mentions as motivation that initially in a washout session you would see a dip in behavior before monkey learns to return to the intuitive mapping; if there is no dip, this might be a good session to look at
- Matt seemed to find with Steve's data that some sessions showed this, others did not
- Look at washout neural activity, feed it through both mappings (use IME), and compare the distance between the decoded velocities, and compare that to if you use intuitive activity
- AND: look at washout data through WMP, and intuitive data through WMP
- Aaron says maybe we wouldn't see this effect with single-day OMPs? Might be something to look for. Or even multi-day OMPs?
- Steve says there's a Shadmehr paper that suggests if he experiences an error in the washout mapping, he might revert to something that would work with the WMP