linear regression version of CaSyn synaptic calcium issues

[rcoreilly]

To normalize sum of CaP weights to same sum as CaD or not?

• for 25 cyc bins, renorm factor is 0.9843 -- not enough by itself to allow CaPScale = 1 to work -- CaPScale ~ 0.95 is generally beneficial
• in lvis, CaPScale with 25 bins, CapScale = 0.95 fails. .96 works. Thus, better to not renormalize, use .95 for this sim.
• lvis fails with bins = 10 -- and most sims are worse with 10, but bgdorsal definitely benefits from 10 cycle bins.
• thus, need to be able to use both 10 and 25 bins and select between them for different projections..
• bins are recorded at the neuron level so actually need to select this at the neuron level.
• need to investigate more about what makes bgdorsal tick with this, in light of updated params, and what makes lvis fail..

Key point: need to use 10 cycle bins always and integrate over them to get 20,30 cycle windows -- can do weighting to get graded effective bin size.. Integration param is per synapse so we can figure out where it matters.

[rcoreilly]

The 10 cycle + integration "envelope" is now all working. By plotting the coefficients I was able to get the function generator code to be essentially identical to final regressed values -- turns out there was some significant CaD error which explained the need for CaPScale < 1. LVis still likes the effective 25msec envelope so far, but 20 learns more quickly initially and then loses it. More exploration underway.

bgdorsal likewise still likes 10msec envelope, but now can explore where that actually matters..

[rcoreilly]

A key feature of the new envelope mechanism is that it provides a sliding window at the 10 msec level, whereas before 25 msec for example imposed a strong aliasing effect for those discrete 25 msec bins. This may explain why bgdorsal actually did better with the 25ms window vs. 10ms. it turns out. And with a slightly higher CaPScale, LVis is now doing better (so far..) at 20ms..

[rcoreilly]

One issue with the envelope timing is that the standard plus phase of 50 msec does not align evenly with the 20 or 30 msec windows -- should try with 60 and 40 cycle plus phases too, adjusting taus accordingly.