Is it possible to somehow run GLM in online mode (as opposed to the batch mode)?
Here's what I mean:
1.Assume there's a segment of SC recording, say from time 0 to time T ('segment 1').
2.Assume that GLM is run on this segment ('glm 1').
3.Now assume a new segment of SC becomes available continuing segment 1 ('segment 2').
Is it possible to run GLM on the combined segment (segments 1 and 2) without changing glm 1, i.e. glm 1 becomes a prefix of the combined glm?
Thx,
David
Last edit: qwer1304 2013-02-09
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
you ask for an iterative algorithm that takes the data to estimate parameters for the first trial, then fixes them and proceeds to the next one, and so on.
This is not possible within the GLM framework (for reasons due to accumulation of errors) but it is precisely what the DCM algorithm for event-related responses does. If you define one "fixed" response per trial, you would get exactly what you need.
We have a paper under review in which we, among other things, compare the variance in single-trial estimates stemming from GLM and DCM approaches. (This was partially inspired by your previous questions). The two approaches are comparable, although the DCM parameter estimates had, across the group, significantly lower residual variance than the GLM estimates.
Hope this helps
Best
Dom
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi,
Is it possible to somehow run GLM in online mode (as opposed to the batch mode)?
Here's what I mean:
1.Assume there's a segment of SC recording, say from time 0 to time T ('segment 1').
2.Assume that GLM is run on this segment ('glm 1').
3.Now assume a new segment of SC becomes available continuing segment 1 ('segment 2').
Is it possible to run GLM on the combined segment (segments 1 and 2) without changing glm 1, i.e. glm 1 becomes a prefix of the combined glm?
Thx,
David
Last edit: qwer1304 2013-02-09
Hi David
you ask for an iterative algorithm that takes the data to estimate parameters for the first trial, then fixes them and proceeds to the next one, and so on.
This is not possible within the GLM framework (for reasons due to accumulation of errors) but it is precisely what the DCM algorithm for event-related responses does. If you define one "fixed" response per trial, you would get exactly what you need.
We have a paper under review in which we, among other things, compare the variance in single-trial estimates stemming from GLM and DCM approaches. (This was partially inspired by your previous questions). The two approaches are comparable, although the DCM parameter estimates had, across the group, significantly lower residual variance than the GLM estimates.
Hope this helps
Best
Dom
Hi Dom,
Thx (as usual) for the TERRIFIC support.
The only issue with DCM approach is that it's computationally SOOOOOO much slower. :(
BTW, since my questions served as an inspiration for the paper, would you mind sharing the draft (no credit is requested :) )
Thx,
David
Hi David
I'd be happy to send it to you if you send me an email with your affiliation (I'll tell you why I'm asking for this in my reply).
Best,
Dom